$IREN Thesis: H2 2026 - 2027 Financial Model:...
Jim Liu@jiahanjimliu
47 views
Jul 22, 2026
~11 min read
Advertisement
1
$IREN Thesis: H2 2026 - 2027
Financial Model: jiahanliu.github.io/IREN-Community/
Previous Thesis Review
I'd like to start each updated thesis by reflecting on what went right and wrong on my previous thesis (1) and if the wrongs are being addressed.
Although I flagged HBM as potential trajectory altering bottleneck in February (2), I saw the HBM bottleneck impact as increased Capex but did not foresee the issues arising from delays in GPU deliveries.
You see, GPUs and HBM are co-packaged. This means, Nvidia secures the HBM supply and then TSMC integrates the GPU and HBM into the same package to make a GPU chip. Nvidia then controls the GPU chip deliveries to hyperscalers or server integrators like Dell, SMCI, and Lenovo.
IREN was getting GPUs from both Dell and Lenovo late which actually means Nvidia did not prioritize them. In ramping up a datacenter, there's only so much theory based preparation IREN can do. In hardware engineering, ramp up problems can arise sequentially. Only once you get the a significant batch of GPUs and run them for longer periods of time, do you uncover cooling deficiencies. Only once you get the cooling right, can you run networking stress test scripts. I'm sure IREN had built in slack time to mitigate delays but there's no mitigation for GPU deliveries arriving months late. Thus we have seen a painful revenue ramp for IREN in H1 2026.
For 2027, the largest change was that Nvidia and IREN formed a strategic partnership to accelerate deployment of AI Infrastructure (3). Contract wise, this is written has Nvidia gets options to buy IREN at $70 vesting upon "Nvidia GPU infrastructure is deployed across IREN campuses and only fully vest upon deployment of 600k GPUs" (6:40-6:55 of 4). However, for Nvidia, these options aren't the true motivator, but rather it's about expanding their ecosystem as I will explain later.
For H2 2026, the change will be that B300/GB300 production will be fully ramped. In 2025-H1 2026, beyond the unprecedented demand, exacerbating the demand-supply imbalance was that the B200/GB200 was a smaller generation of GPUs compared to H100/H200 and B300/GB300 because the Blackwell ramp had faced technical and supply chain challenges (5).
Demand Review
In early February, I identified that Anthropic would release unprecedented growth numbers would result in urgent GPU demand (6). While the demand from Anthropic has played out, Anthropic subsequently signed with everyone possible from CRWV (7), xAI to Akamai Cloud (8) and Amazon, Google, Microsoft. IREN opted to make their flagship SW1 campus be Vera Rubins rather than GB300s but I will explain why this is a prime site for Anthropic.
The Core Thesis
Every AI thesis should be firmly grounded on roadmap and incentives of the AI Hyperscalers: Nvidia, Anthropic and OpenAI.
The previous generation of Hyperscalers followed the Amazon model: excel at cloud infrastructure for in-house projects and provide managed software services for enterprises applications.
The AI generation of Hyperscalers will follow the Anthropic model: excel at agentic AI that builds software in house and provide agentic AI for enterprises to build vertically integrated applications and tailored software infrastructure.
Why vertically integrated applications and tailored software infrastructure? Contrary to misconception, AI does not make SaaS obsolete but rather raises the bar for SaaS. Just like how excelling at Leetcode and reading DDA is no longer sufficient for software engineering interviews, application level software is no longer sufficient for SaaS companies.
We saw the following business solidify their moat through vertically integrated applications and tailored software infrastructure:
2000s: Amazon, Google
2010-2025: Meta, Netflix, Uber, Salesforce, Airbnb, TikTok, Palantir, Tesla.
In the 2026-2030 AI will enable smaller teams to output more software and having vertically integrated applications and tailored software infrastructure will be a requirement of SaaS companies to have proprietary services and squeeze out optimizations. In other words: if you've played around with Claude Code at all, you will know that application level software by itself is not a differentiated business.
Beyond the OpenAI and Anthropic, there will be a important role for open source and custom models. However, generating application code and connecting it to Token APIs will not be a differentiated business. The margins will come from optimizing the inference stack based on application call patterns down to bare metal. Some companies like Cursor and $TEM have gone as far as building their own custom models to derive proprietary differentiation and accrue margins.
Open Source Models
Open Source will be important. Open Source Models will be like Linux: very important but optimizing Linux is not a big business. Many leading enterprises have in-house customized distributions of Linux optimized for their workloads.
While there is alot of chatter about how Anthropic and OpenAI Token cost have become borderline untenable (9), the thing you have to understand is that Anthropic and OpenAI are building a ecosystem not a token generator. If Open Source takes significant enterprise market share because it's cheaper, Anthropic and OpenAI will have different tier models on the cost curve. Both Anthropic and OpenAI need high market share to achieve economics of scale and form an ecosystem. They already have cost tiered models but if Open Source starts to take significant enterprise market share, OpenAI and Anthropic will get more aggressive.
Let's me put it another way: principal engineers at Google are using Claude Code to build GCP (10). Claude Code is improving rapidly. Do you think the AI Natives and Leading Enterprises of tomorrow will build their own software infrastructure or pay out 80% margins to Neoclouds?
IREN's Role
I see IREN's software acquisitions in Mirantis as a 2-3 year stop gap while AI Natives and Leading Enterprises ramp up on AI while Claude Code continues to improve.
In the age where Nvidia/Anthropic/OpenAI are the Hyperscalers, I see IREN as the Exxon ($XOM). XOM does alot more than producing oil and gas, they do all the downstream work to refine, distribute and create chemcial dervatives of oil. Likewise, IREN has the expertise to do power studies to secure grid connected power and build out power infrastructure but vertically integrated in the sense it does datacenter design, datacenter operations for hardware uptime, and managed Kubernetes.
Some who are trying to invest in Neoclouds as an AI play are ignoring the risks of how AI will disrupt software. AI does not make software obsolete at all but it increases the software which leading enterprise will have optimized in-house. There will be many successful Colocation providers and Neoclouds but IREN unique in that it has more vertical integration than colocation providers (CIFR, WULF, HUT) but without higher valuation premium for the software layer (CRWV, NBIS) primed for disruption and more grid connected power secured than anyone else outside the old HS.
Other Neocloud investors may say electricity is cheap but once you look $BE, you realize secured power is valuable. The margin which other Neoclouds like $ORCL and $NBIS give up to BE is margin advantage for $IREN. In other words, with 4.9GW of secured power and multi-GW pipeline, $BE and $IREN have a shared-factor exposure to power but with $BE market cap at 86B, IREN's power component isn't properly valued yet.
Anthropic/OpenAI as Foretellers of Software in the Age of AI
Just how Amazon pioneer web architecture which became the forerunner runner of managed services, Anthropic and OPenAI are the pioneers are AI driven software development and serve as guidance of how the software landscape will develop in the next 5 years.
Those who think Anthropic is depending on AWS, CRWV or god forbid Akamai for Cloud Infrastructure don't understand that Anthropic is the best software company on earth. Prior that title belong to Google who developed all their software infrasturcture in house. Anthropic was recently hiring for engineers for ROCm (11), this shows that Anthropic is developing their entire inference stack down the AMD GPU. You bet that already have optimized inference and training stacks for Nvidia GPUs if they are already working on porting it to AMU GPUs.
OpenAI is partnering with Dell to bring Codex to on-prem (12). On-prem means deploying Codex to an enterprise's own datacenter. Clearly this means OpenAI has their own inference stack. There's no reason for OpenAI to make this an Dell exclusive, OpenAI will allow enterprises to deploy Codex onto bare metal as this allows them to expand their market share beyond their allocated compute. This will be a tool as they fight Anthropic and Open Source for market share.
The takeaway is that Anthropic and OpenAI have optimize their model, inference stack, and workload orchestration all the way to the accelerator. Anthropic and OpenAI will be running optimally whether on CRWV, NBIS, or IREN bare metal GPUs. No matter if you are on CRWV, NBIS or IREN bare metal, Nvidia instruction set architecture is the same. For Anthropic who brings their own inference stack, none of the Neocloud software matters.
Nvidia
Now why would Nvidia be incentivize to increase IREN's priority for GPU deliveries? @Agrippa_Inv and @franklee6924T have written extensively about NVIDIA DSX initiative where IREN's SW1 site will the "flagship deployment for Nvidia's DSX architecture but I'll explain it from a historical angle.
Nvidia has extremely well in building a moat from iterating on CUDA, to buying Mellanox to dominate backend inter-rack GPU networking, to buying Groq for low latency inference. However Nvidia has a key risk as long as AWS, Azure, GCP own the customer relationship, the risk of these Hyperscalers developing their own ASIC is always there. Granted these ASICs are not immediately threatening to Nvidia, Jensen works on long foresight and prevents threats before they rise.
Jensen practically brought up CRWV to hedge against the Hyperscalers and then gave strong backing to NBIS. Now, it's strategic parternship with IREN is the third leg to hedge against the trio AWS, Azure and GCP.
In the 1990s, Wintel (Windows Intel) dominated the margins. Echoing Andy Grove's strategy of commoditizing your complement, Nvidia is trying to commoditize the current hypperscalers. IREN might yet be the best match for Nvidia strategy because it's about monetizing power at scale and not trying to grow margin on the inference stack. In other words, Nvidia is the modern day bigger Intel, Anthropic/OpenAI are the modern day bigger Windows/OS X, and IREN, CRWV, NBIS are the modern day bigger Dell, IBM, Compaq. In the page of AI, infrastructure buildout will be much larger than PC integrators and IREN will be XOM scale DC/IaaS buildout.
Research Posts
Above is the overarching view on the IREN thesis. I have written posts covering individual components of the IREN thesis and will continue to cover developments.
Financial Model: x.com/jiahanjimliu/s…
Power Bottleneck: x.com/jiahanjimliu/s…
Value of Secured Power: x.com/jiahanjimliu/s…
IaaS and PaaS Markets: x.com/jiahanjimliu/s…
Why AI Research Breakthrough that drastically Reduces Need for GPUs is Highly Unlikely: x.com/jiahanjimliu/s…
Edge AI: x.com/jiahanjimliu/s…
Netflix Case Study: x.com/jiahanjimliu/s…
Open Source Models on Open Sourced Infrastructure: x.com/jiahanjimliu/s…
Mirantis Strategy: x.com/jiahanjimliu/s…
Mirantis Technical Capabilities: x.com/jiahanjimliu/s…
Mirantis Sovereign AI: x.com/jiahanjimliu/s…
Datacenter Vertical Integration: x.com/jiahanjimliu/s…
Hybrid DLC + RDHx Cooling:
Credits
X accounts actively posting IREN Research that I read:
OGs who research I've read from $5:
@FransBakker9812 - analyzes everything IREN including satellite images, documents for powered land developments, and site employee hearsay for his sub group
@Agrippa_Inv - the cleanest thesis and long form research
@_Sgr_A_Star - gets deep into financial releases
@bitcoinbutcher1 - sunday spaces lead
@Umbisam - risk cautious but not risk adverse insights
@nanotitan28 - doing IREN TA since day 1
Large Accounts with Excellent IREN Coverage
@TheTechInvest - great coverage of Tech Stocks with all in IREN allocation, previously all-in Nvidia
@kevinxu - 8 figure successful investor with high IREN conviction
@moninvestor - small/mid cap specialist who fully understands the IREN thesis
Industry Coverage:
@scludweed/@alanbialo - great repost and who happen to be whales. Not listed here but the biggest retail whale has a bigger allocation than seed investors but shouldn't be revealed for privacy reasons.
@MarkosAAIG - Day 3 NBIS Investor Industry Coverage
@pepe_maltese - Institutional Grade POV
@GlobalCollapse - options dealing
@XCapitalMgmt - the legit account covering IREN with Capital in it's name
@GyujinAAIG - Korean Medical Resident also covering IREN, Semis and Materials
@ilzmcfly - forensic grade digging
@StockAnalystPro - AI Director at MSFT POV
Seed Investors: @BTCYESPLS, @mikealfred, @TheBigDegen, @roberto45580514
Financial Model: jiahanliu.github.io/IREN-Community/
Previous Thesis Review
I'd like to start each updated thesis by reflecting on what went right and wrong on my previous thesis (1) and if the wrongs are being addressed.
Although I flagged HBM as potential trajectory altering bottleneck in February (2), I saw the HBM bottleneck impact as increased Capex but did not foresee the issues arising from delays in GPU deliveries.
You see, GPUs and HBM are co-packaged. This means, Nvidia secures the HBM supply and then TSMC integrates the GPU and HBM into the same package to make a GPU chip. Nvidia then controls the GPU chip deliveries to hyperscalers or server integrators like Dell, SMCI, and Lenovo.
IREN was getting GPUs from both Dell and Lenovo late which actually means Nvidia did not prioritize them. In ramping up a datacenter, there's only so much theory based preparation IREN can do. In hardware engineering, ramp up problems can arise sequentially. Only once you get the a significant batch of GPUs and run them for longer periods of time, do you uncover cooling deficiencies. Only once you get the cooling right, can you run networking stress test scripts. I'm sure IREN had built in slack time to mitigate delays but there's no mitigation for GPU deliveries arriving months late. Thus we have seen a painful revenue ramp for IREN in H1 2026.
For 2027, the largest change was that Nvidia and IREN formed a strategic partnership to accelerate deployment of AI Infrastructure (3). Contract wise, this is written has Nvidia gets options to buy IREN at $70 vesting upon "Nvidia GPU infrastructure is deployed across IREN campuses and only fully vest upon deployment of 600k GPUs" (6:40-6:55 of 4). However, for Nvidia, these options aren't the true motivator, but rather it's about expanding their ecosystem as I will explain later.
For H2 2026, the change will be that B300/GB300 production will be fully ramped. In 2025-H1 2026, beyond the unprecedented demand, exacerbating the demand-supply imbalance was that the B200/GB200 was a smaller generation of GPUs compared to H100/H200 and B300/GB300 because the Blackwell ramp had faced technical and supply chain challenges (5).
Demand Review
In early February, I identified that Anthropic would release unprecedented growth numbers would result in urgent GPU demand (6). While the demand from Anthropic has played out, Anthropic subsequently signed with everyone possible from CRWV (7), xAI to Akamai Cloud (8) and Amazon, Google, Microsoft. IREN opted to make their flagship SW1 campus be Vera Rubins rather than GB300s but I will explain why this is a prime site for Anthropic.
The Core Thesis
Every AI thesis should be firmly grounded on roadmap and incentives of the AI Hyperscalers: Nvidia, Anthropic and OpenAI.
The previous generation of Hyperscalers followed the Amazon model: excel at cloud infrastructure for in-house projects and provide managed software services for enterprises applications.
The AI generation of Hyperscalers will follow the Anthropic model: excel at agentic AI that builds software in house and provide agentic AI for enterprises to build vertically integrated applications and tailored software infrastructure.
Why vertically integrated applications and tailored software infrastructure? Contrary to misconception, AI does not make SaaS obsolete but rather raises the bar for SaaS. Just like how excelling at Leetcode and reading DDA is no longer sufficient for software engineering interviews, application level software is no longer sufficient for SaaS companies.
We saw the following business solidify their moat through vertically integrated applications and tailored software infrastructure:
2000s: Amazon, Google
2010-2025: Meta, Netflix, Uber, Salesforce, Airbnb, TikTok, Palantir, Tesla.
In the 2026-2030 AI will enable smaller teams to output more software and having vertically integrated applications and tailored software infrastructure will be a requirement of SaaS companies to have proprietary services and squeeze out optimizations. In other words: if you've played around with Claude Code at all, you will know that application level software by itself is not a differentiated business.
Beyond the OpenAI and Anthropic, there will be a important role for open source and custom models. However, generating application code and connecting it to Token APIs will not be a differentiated business. The margins will come from optimizing the inference stack based on application call patterns down to bare metal. Some companies like Cursor and $TEM have gone as far as building their own custom models to derive proprietary differentiation and accrue margins.
Open Source Models
Open Source will be important. Open Source Models will be like Linux: very important but optimizing Linux is not a big business. Many leading enterprises have in-house customized distributions of Linux optimized for their workloads.
While there is alot of chatter about how Anthropic and OpenAI Token cost have become borderline untenable (9), the thing you have to understand is that Anthropic and OpenAI are building a ecosystem not a token generator. If Open Source takes significant enterprise market share because it's cheaper, Anthropic and OpenAI will have different tier models on the cost curve. Both Anthropic and OpenAI need high market share to achieve economics of scale and form an ecosystem. They already have cost tiered models but if Open Source starts to take significant enterprise market share, OpenAI and Anthropic will get more aggressive.
Let's me put it another way: principal engineers at Google are using Claude Code to build GCP (10). Claude Code is improving rapidly. Do you think the AI Natives and Leading Enterprises of tomorrow will build their own software infrastructure or pay out 80% margins to Neoclouds?
IREN's Role
I see IREN's software acquisitions in Mirantis as a 2-3 year stop gap while AI Natives and Leading Enterprises ramp up on AI while Claude Code continues to improve.
In the age where Nvidia/Anthropic/OpenAI are the Hyperscalers, I see IREN as the Exxon ($XOM). XOM does alot more than producing oil and gas, they do all the downstream work to refine, distribute and create chemcial dervatives of oil. Likewise, IREN has the expertise to do power studies to secure grid connected power and build out power infrastructure but vertically integrated in the sense it does datacenter design, datacenter operations for hardware uptime, and managed Kubernetes.
Some who are trying to invest in Neoclouds as an AI play are ignoring the risks of how AI will disrupt software. AI does not make software obsolete at all but it increases the software which leading enterprise will have optimized in-house. There will be many successful Colocation providers and Neoclouds but IREN unique in that it has more vertical integration than colocation providers (CIFR, WULF, HUT) but without higher valuation premium for the software layer (CRWV, NBIS) primed for disruption and more grid connected power secured than anyone else outside the old HS.
Other Neocloud investors may say electricity is cheap but once you look $BE, you realize secured power is valuable. The margin which other Neoclouds like $ORCL and $NBIS give up to BE is margin advantage for $IREN. In other words, with 4.9GW of secured power and multi-GW pipeline, $BE and $IREN have a shared-factor exposure to power but with $BE market cap at 86B, IREN's power component isn't properly valued yet.
Anthropic/OpenAI as Foretellers of Software in the Age of AI
Just how Amazon pioneer web architecture which became the forerunner runner of managed services, Anthropic and OPenAI are the pioneers are AI driven software development and serve as guidance of how the software landscape will develop in the next 5 years.
Those who think Anthropic is depending on AWS, CRWV or god forbid Akamai for Cloud Infrastructure don't understand that Anthropic is the best software company on earth. Prior that title belong to Google who developed all their software infrasturcture in house. Anthropic was recently hiring for engineers for ROCm (11), this shows that Anthropic is developing their entire inference stack down the AMD GPU. You bet that already have optimized inference and training stacks for Nvidia GPUs if they are already working on porting it to AMU GPUs.
OpenAI is partnering with Dell to bring Codex to on-prem (12). On-prem means deploying Codex to an enterprise's own datacenter. Clearly this means OpenAI has their own inference stack. There's no reason for OpenAI to make this an Dell exclusive, OpenAI will allow enterprises to deploy Codex onto bare metal as this allows them to expand their market share beyond their allocated compute. This will be a tool as they fight Anthropic and Open Source for market share.
The takeaway is that Anthropic and OpenAI have optimize their model, inference stack, and workload orchestration all the way to the accelerator. Anthropic and OpenAI will be running optimally whether on CRWV, NBIS, or IREN bare metal GPUs. No matter if you are on CRWV, NBIS or IREN bare metal, Nvidia instruction set architecture is the same. For Anthropic who brings their own inference stack, none of the Neocloud software matters.
Nvidia
Now why would Nvidia be incentivize to increase IREN's priority for GPU deliveries? @Agrippa_Inv and @franklee6924T have written extensively about NVIDIA DSX initiative where IREN's SW1 site will the "flagship deployment for Nvidia's DSX architecture but I'll explain it from a historical angle.
Nvidia has extremely well in building a moat from iterating on CUDA, to buying Mellanox to dominate backend inter-rack GPU networking, to buying Groq for low latency inference. However Nvidia has a key risk as long as AWS, Azure, GCP own the customer relationship, the risk of these Hyperscalers developing their own ASIC is always there. Granted these ASICs are not immediately threatening to Nvidia, Jensen works on long foresight and prevents threats before they rise.
Jensen practically brought up CRWV to hedge against the Hyperscalers and then gave strong backing to NBIS. Now, it's strategic parternship with IREN is the third leg to hedge against the trio AWS, Azure and GCP.
In the 1990s, Wintel (Windows Intel) dominated the margins. Echoing Andy Grove's strategy of commoditizing your complement, Nvidia is trying to commoditize the current hypperscalers. IREN might yet be the best match for Nvidia strategy because it's about monetizing power at scale and not trying to grow margin on the inference stack. In other words, Nvidia is the modern day bigger Intel, Anthropic/OpenAI are the modern day bigger Windows/OS X, and IREN, CRWV, NBIS are the modern day bigger Dell, IBM, Compaq. In the page of AI, infrastructure buildout will be much larger than PC integrators and IREN will be XOM scale DC/IaaS buildout.
Research Posts
Above is the overarching view on the IREN thesis. I have written posts covering individual components of the IREN thesis and will continue to cover developments.
Financial Model: x.com/jiahanjimliu/s…
Power Bottleneck: x.com/jiahanjimliu/s…
Value of Secured Power: x.com/jiahanjimliu/s…
IaaS and PaaS Markets: x.com/jiahanjimliu/s…
Why AI Research Breakthrough that drastically Reduces Need for GPUs is Highly Unlikely: x.com/jiahanjimliu/s…
Edge AI: x.com/jiahanjimliu/s…
Netflix Case Study: x.com/jiahanjimliu/s…
Open Source Models on Open Sourced Infrastructure: x.com/jiahanjimliu/s…
Mirantis Strategy: x.com/jiahanjimliu/s…
Mirantis Technical Capabilities: x.com/jiahanjimliu/s…
Mirantis Sovereign AI: x.com/jiahanjimliu/s…
Datacenter Vertical Integration: x.com/jiahanjimliu/s…
Hybrid DLC + RDHx Cooling:
Credits
X accounts actively posting IREN Research that I read:
OGs who research I've read from $5:
@FransBakker9812 - analyzes everything IREN including satellite images, documents for powered land developments, and site employee hearsay for his sub group
@Agrippa_Inv - the cleanest thesis and long form research
@_Sgr_A_Star - gets deep into financial releases
@bitcoinbutcher1 - sunday spaces lead
@Umbisam - risk cautious but not risk adverse insights
@nanotitan28 - doing IREN TA since day 1
Large Accounts with Excellent IREN Coverage
@TheTechInvest - great coverage of Tech Stocks with all in IREN allocation, previously all-in Nvidia
@kevinxu - 8 figure successful investor with high IREN conviction
@moninvestor - small/mid cap specialist who fully understands the IREN thesis
Industry Coverage:
@scludweed/@alanbialo - great repost and who happen to be whales. Not listed here but the biggest retail whale has a bigger allocation than seed investors but shouldn't be revealed for privacy reasons.
@MarkosAAIG - Day 3 NBIS Investor Industry Coverage
@pepe_maltese - Institutional Grade POV
@GlobalCollapse - options dealing
@XCapitalMgmt - the legit account covering IREN with Capital in it's name
@GyujinAAIG - Korean Medical Resident also covering IREN, Semis and Materials
@ilzmcfly - forensic grade digging
@StockAnalystPro - AI Director at MSFT POV
Seed Investors: @BTCYESPLS, @mikealfred, @TheBigDegen, @roberto45580514
Jim Liu@jiahanjimliu
$IREN Hybrid DLC + RDHx Cooling
@CernunnosCap posts can a good place to find misconceptions that to be clarified. Here he states that Horizon 1 GB300 is only using Motivair RDHx cooling.
However, Hybrid DLC + RDHx Cooling is known to people who actually work in the industry (1). DLC (Direct Liquid Cooling) cools the higher heat GPUs and CPUs. RDHx (Rear Door Heat Exchange) cools the Networking, Management, Storage equipment which has lower cooling requirements.
The design tradeoff is that cold plates have very tiny channels and the to create enough flow volume, the dense cooling fluid needs to move very fast which means high pressure on the pipes. These is the steel pipes you see in the H1 liquid cooling picture. High pressure means more maintenance and operational monitoring overhead to prevent leaks.
Thus for components like Networking, Management, Storage which produce way less heat than the GPUs and CPUs, it's let RDHx do these components because RDHx has lower operational complexity. These are the flexible non-steel pipes you see in the H1 liquid cooling picture and the Motivair sample picture that @CernunnosCap provided.
Any serious engineer can see that $IREN is minimizing operational complexity and maximizing uptime with Hybrid DLC + RDHx Cooling. However to the untrained eye, it's easy to cherry-pick one component and not recognize the full system about it.
(1) eliovp.com/building-the-e…
@CernunnosCap posts can a good place to find misconceptions that to be clarified. Here he states that Horizon 1 GB300 is only using Motivair RDHx cooling.
However, Hybrid DLC + RDHx Cooling is known to people who actually work in the industry (1). DLC (Direct Liquid Cooling) cools the higher heat GPUs and CPUs. RDHx (Rear Door Heat Exchange) cools the Networking, Management, Storage equipment which has lower cooling requirements.
The design tradeoff is that cold plates have very tiny channels and the to create enough flow volume, the dense cooling fluid needs to move very fast which means high pressure on the pipes. These is the steel pipes you see in the H1 liquid cooling picture. High pressure means more maintenance and operational monitoring overhead to prevent leaks.
Thus for components like Networking, Management, Storage which produce way less heat than the GPUs and CPUs, it's let RDHx do these components because RDHx has lower operational complexity. These are the flexible non-steel pipes you see in the H1 liquid cooling picture and the Motivair sample picture that @CernunnosCap provided.
Any serious engineer can see that $IREN is minimizing operational complexity and maximizing uptime with Hybrid DLC + RDHx Cooling. However to the untrained eye, it's easy to cherry-pick one component and not recognize the full system about it.
(1) eliovp.com/building-the-e…
2
(1) x.com/jiahanjimliu/s…
(2) x.com/jiahanjimliu/s…
(3) nvidianews.nvidia.com/news/nvidia-an…
(4) edge.media-server.com/mmc/p/8oxdtymr/
(5) datacenterdynamics.com/en/news/nvidia…
(6) x.com/jiahanjimliu/s…
(7) x.com/jiahanjimliu/s…
(8) x.com/jiahanjimliu/s…
(9)
(10) x.com/jiahanjimliu/s…
(11) x.com/jiahanjimliu/s…
(12) x.com/jiahanjimliu/s…
(13) x.com/Agrippa_Inv/st…
(2) x.com/jiahanjimliu/s…
(3) nvidianews.nvidia.com/news/nvidia-an…
(4) edge.media-server.com/mmc/p/8oxdtymr/
(5) datacenterdynamics.com/en/news/nvidia…
(6) x.com/jiahanjimliu/s…
(7) x.com/jiahanjimliu/s…
(8) x.com/jiahanjimliu/s…
(9)
(10) x.com/jiahanjimliu/s…
(11) x.com/jiahanjimliu/s…
(12) x.com/jiahanjimliu/s…
(13) x.com/Agrippa_Inv/st…
Hedgie@HedgieMarkets
🦔Fortune published a piece this afternoon connecting Microsoft and Uber's AI cost overruns to token economics, with a headline that lands hard: "Microsoft reports are exposing AI's real cost problem: Using the tech is more expensive than paying human employees." Underneath those headlines, the unit economics tell the story. OpenAI is projected to lose $14 billion in 2026, spending roughly $2 for every dollar of revenue it brings in. Anthropic is in a similar position with break-even not projected until 2028. GPU rental prices for Nvidia's newest Blackwell chips jumped 48% in just two months. OpenAI's response was to close a $122 billion private funding round at an $852 billion valuation, the largest in history.
My Take
The token pricing story is really an IPO timing story. OpenAI, Anthropic, and xAI all need to go public in the next 18 to 24 months because the private market cannot keep absorbing burn rates like these indefinitely. Public markets do not accept "we will figure it out" as a line item on an S-1, they require disclosed unit economics with a credible path to profitability and a date attached. That deadline is why the price increases are happening now rather than next year. The labs need to show declining loss curves before the filings hit, and that means enterprise customers have to start covering more of the actual cost regardless of whether the productivity math holds on their end.
Every token bought over the last two years was effectively subsidized below cost by venture capital and hyperscaler cross-subsidies, and that subsidy has a hard deadline. Uber publicly admitted burning through its entire 2026 AI budget in four months, and CFOs at major enterprises are starting to flag the same pressure. The labs cannot keep losing $2 per dollar of revenue once they file public statements, so the cost transfer to customers accelerates from here. For investors, the question is not whether these companies are valuable. They clearly are. The question is who absorbs the difference between what enterprises can budget and what the models actually consume between now and 2028, and right now the answer is the hyperscalers funding the buildout. That is why I have been watching Microsoft and Amazon capex commentary more closely than the lab announcements themselves.
Hedgie🤗
Link: fortune.com/2026/05/22/mic…
My Take
The token pricing story is really an IPO timing story. OpenAI, Anthropic, and xAI all need to go public in the next 18 to 24 months because the private market cannot keep absorbing burn rates like these indefinitely. Public markets do not accept "we will figure it out" as a line item on an S-1, they require disclosed unit economics with a credible path to profitability and a date attached. That deadline is why the price increases are happening now rather than next year. The labs need to show declining loss curves before the filings hit, and that means enterprise customers have to start covering more of the actual cost regardless of whether the productivity math holds on their end.
Every token bought over the last two years was effectively subsidized below cost by venture capital and hyperscaler cross-subsidies, and that subsidy has a hard deadline. Uber publicly admitted burning through its entire 2026 AI budget in four months, and CFOs at major enterprises are starting to flag the same pressure. The labs cannot keep losing $2 per dollar of revenue once they file public statements, so the cost transfer to customers accelerates from here. For investors, the question is not whether these companies are valuable. They clearly are. The question is who absorbs the difference between what enterprises can budget and what the models actually consume between now and 2028, and right now the answer is the hyperscalers funding the buildout. That is why I have been watching Microsoft and Amazon capex commentary more closely than the lab announcements themselves.
Hedgie🤗
Link: fortune.com/2026/05/22/mic…
3
Potential Customer POV: How Anthropic Plans Compute Needs:
Jim Liu@jiahanjimliu
Anthropic CFO Krishna Rao Interview: How They Plan Compute
youtube.com/watch?v=wEEZPp…
1. They buy Nvidia GPUs, Google TPUs, Amazon Trainium and use all 3 fungibly.
2. Built in-house orchestration to utilize each generation of different chip most efficiently.
3. Works as close to the bare metal on all 3 chips as possible to have maximum flexiblity and efficiency.
4. Has influence over Amazon Trainium design roadmap. Sometimes works directly with Annapurna engineers.
5. They have allocation for their own engineers separate from research which in itself has an allocation.
6. Model Development has highest priority even over Customer allocation.
7. ROI of being on the best frontier model is very high, especially enterprise. Immediately when Opus 4.7 comes out, almost everyone switches to it even though it's more expensive.
8. Buying bare metal allows them have granular real-time flexibility on re-allocating compute between research (model development), internal engineering, customers.
9. Their metric for how good a model is based on statistic collected from customer not benchmarks.
10. Sometimes it's not Anthropic that pushes the frontier. Customers unlock TAM beyond what Anthropic builds the model for. Then Anthropic push that frontier.
11. Scaling laws are alive and well. 90%+ of internal code is written by Claude Code including Claude Code itself.
12. Near-term compute: will buy wherever they can get it.
13. Signed up to 5GW from for Google TPU and Amazon Trainium. These ASICs are not as good as Nvidia GPUs but this is what keeps Jensen up at night.
14. Thinks of Anthropic as a Platform like early days of AWS. Like I said, Anthropic is the AI age hyperscaler.
15. Prices tokens to be accessible to gain market share and at the same time go towards cashflow positive.
16. Claude Code is used as a Coworker on the Finance Teams. All financial statements used with Claude and human checks it. Claude also does analysis and helps drive decision.
Jim's Commentary
I know $IREN just acquired Mirantis but I still believe bare metal is the biggest business line by volume. Anthropic CFO just showed the advantages of buying metal and then in-housing all the software.
Some people may not know but
1. Amazon Bedrock = SaaS
2. Amazon SageMaker = PaaS
3. Amazon EC2 P5, P6, P4, P3 = Bare Metal IaaS
Not publically disclosed but anyone working as a software engineer would know that by revenue, EC2 P Series is Amazon's biggest AI Cloud segment.
youtube.com/watch?v=wEEZPp…
1. They buy Nvidia GPUs, Google TPUs, Amazon Trainium and use all 3 fungibly.
2. Built in-house orchestration to utilize each generation of different chip most efficiently.
3. Works as close to the bare metal on all 3 chips as possible to have maximum flexiblity and efficiency.
4. Has influence over Amazon Trainium design roadmap. Sometimes works directly with Annapurna engineers.
5. They have allocation for their own engineers separate from research which in itself has an allocation.
6. Model Development has highest priority even over Customer allocation.
7. ROI of being on the best frontier model is very high, especially enterprise. Immediately when Opus 4.7 comes out, almost everyone switches to it even though it's more expensive.
8. Buying bare metal allows them have granular real-time flexibility on re-allocating compute between research (model development), internal engineering, customers.
9. Their metric for how good a model is based on statistic collected from customer not benchmarks.
10. Sometimes it's not Anthropic that pushes the frontier. Customers unlock TAM beyond what Anthropic builds the model for. Then Anthropic push that frontier.
11. Scaling laws are alive and well. 90%+ of internal code is written by Claude Code including Claude Code itself.
12. Near-term compute: will buy wherever they can get it.
13. Signed up to 5GW from for Google TPU and Amazon Trainium. These ASICs are not as good as Nvidia GPUs but this is what keeps Jensen up at night.
14. Thinks of Anthropic as a Platform like early days of AWS. Like I said, Anthropic is the AI age hyperscaler.
15. Prices tokens to be accessible to gain market share and at the same time go towards cashflow positive.
16. Claude Code is used as a Coworker on the Finance Teams. All financial statements used with Claude and human checks it. Claude also does analysis and helps drive decision.
Jim's Commentary
I know $IREN just acquired Mirantis but I still believe bare metal is the biggest business line by volume. Anthropic CFO just showed the advantages of buying metal and then in-housing all the software.
Some people may not know but
1. Amazon Bedrock = SaaS
2. Amazon SageMaker = PaaS
3. Amazon EC2 P5, P6, P4, P3 = Bare Metal IaaS
Not publically disclosed but anyone working as a software engineer would know that by revenue, EC2 P Series is Amazon's biggest AI Cloud segment.
4
Some bear factors to watch out for from @GavinSBaker who I take very seriously:
Jim Liu@jiahanjimliu
@GavinSBaker Interview on AI Buildout
youtube.com/watch?v=Mmj_G9…
Gavin Baker is the OG Leopold Aschenbrenner, one of the best tech investors with connections in Silicon Valley, Defense Tech and Taiwan Fabs. He see's many things before X see's it. He shares many observations that differ from those in the X silo.
1. Space datacenters are coming in 2027-2028. SpaceX will dominate. Training with higher bandwidth requirements between racks will still be done on earth but inference can be done in space. He's spent considerable time at Starbase and he's confident they will go to market in 1-2 years. Earth DCs will still be valuable because the world will need all the compute it can get.
2. TSMC fab shortage is preventing AI from being a bubble but companies are slowly going to Intel and Samsung. TSMC needs to maintain a lead over Intel and Samsung to keep an overbuild from happening.
3. Frontier models capture the most value. $20 - $200 plans all have quantized and pruned CoT models. Only API pricing gets you full powered model nowadays because of compute shortage.
4. The ability to split prefill and decode will increase the useful life of GPUs to 10-15 years. This is because you can use older GPus for pre-fill and have an specialized ASIC to handle decode.
5. Very few ASIC startups will survive. Cerebras is one of the few doing something different and hard but if their market share gets more than 1-2%, Nvidia will come after them and they will not be able to grow to scale. TPU and Trainium are the competitive ASICs
6. Chinese Open Source is stays close to frontier because they distill from American Labs. Once American Labs can prevent Chinese Labs from distilling, it's over. Without distilling, China doesn't have the compute to be on the frontier.
7. Cybersecurity will be highly disrupted. Huge opportunities will be ahead but many security companies today will not be relevant in the age of AI.
8. Shortages will make mediocre businesses look like graet businesses. He directly points to random parts of the AI components that traditionally had single digit margins and are now being pumped on X and Reddit. These companies will have margins for a few years and look like have incredible margins but when supply increases, these stocks will tank.
He doesn't give smaller companies by name but he misses a time when there were more memory bears. Memory is valuable but if supply ever alleviates, 80% margins will not preserve.
The cycle is shortage begins -> margins spike -> Investors treat new margins as normal -> supply responds -> margins collapse. Anyone in commodity investing knows this. There is good research showing HBM has a moat and is way more complex and is not prone to this but we do need intelligent bear cases to test this.
youtube.com/watch?v=Mmj_G9…
Gavin Baker is the OG Leopold Aschenbrenner, one of the best tech investors with connections in Silicon Valley, Defense Tech and Taiwan Fabs. He see's many things before X see's it. He shares many observations that differ from those in the X silo.
1. Space datacenters are coming in 2027-2028. SpaceX will dominate. Training with higher bandwidth requirements between racks will still be done on earth but inference can be done in space. He's spent considerable time at Starbase and he's confident they will go to market in 1-2 years. Earth DCs will still be valuable because the world will need all the compute it can get.
2. TSMC fab shortage is preventing AI from being a bubble but companies are slowly going to Intel and Samsung. TSMC needs to maintain a lead over Intel and Samsung to keep an overbuild from happening.
3. Frontier models capture the most value. $20 - $200 plans all have quantized and pruned CoT models. Only API pricing gets you full powered model nowadays because of compute shortage.
4. The ability to split prefill and decode will increase the useful life of GPUs to 10-15 years. This is because you can use older GPus for pre-fill and have an specialized ASIC to handle decode.
5. Very few ASIC startups will survive. Cerebras is one of the few doing something different and hard but if their market share gets more than 1-2%, Nvidia will come after them and they will not be able to grow to scale. TPU and Trainium are the competitive ASICs
6. Chinese Open Source is stays close to frontier because they distill from American Labs. Once American Labs can prevent Chinese Labs from distilling, it's over. Without distilling, China doesn't have the compute to be on the frontier.
7. Cybersecurity will be highly disrupted. Huge opportunities will be ahead but many security companies today will not be relevant in the age of AI.
8. Shortages will make mediocre businesses look like graet businesses. He directly points to random parts of the AI components that traditionally had single digit margins and are now being pumped on X and Reddit. These companies will have margins for a few years and look like have incredible margins but when supply increases, these stocks will tank.
He doesn't give smaller companies by name but he misses a time when there were more memory bears. Memory is valuable but if supply ever alleviates, 80% margins will not preserve.
The cycle is shortage begins -> margins spike -> Investors treat new margins as normal -> supply responds -> margins collapse. Anyone in commodity investing knows this. There is good research showing HBM has a moat and is way more complex and is not prone to this but we do need intelligent bear cases to test this.
5
@GavinSBaker What to Monitor for 2026:
Jim Liu@jiahanjimliu
$IREN: What to Monitor in 2026
Revenue
No doubt $IREN is rich in power, what $IREN investors need to focus on is revenue. The demand is certainly there, it's about getting GPUs online. Even the rate at which $IREN can sign contract is bottleneck by getting GPUs online because getting GPUs online derisk your ability to meet timelines to sign the next contract.
So what you see in $CRWV is that at the beginning they were really fast and then started slowing down. At first, a Neocloud will 2x revenue every quarter but then the ramp are bounded by the physical world. The benefit of having alot of power is that you'll be able to keep growing at a high rate until you ran out of power or your colocation provider(s) hits their limits. However, that's not $IREN's challenge. $IREN needs to get it's ramp -> cashflow -> ramp feedback loop going. Getting slow GPUs slowed down the whole ramp as you are doing theory based preparation for your datacenter until you get the GPUs and then problems can be uncovered sequentially.
Benchmarks
2024 was $CRWV's ramp year:
Q4 2023: 116m
Q1: 188.7m (+62.7% QoQ)
Q2: 395.4m (+109.5% QoQ)
Q3: 583.9m (+47.7% QoQ)
Q4: 747m (+27.9% QoQ)
Q1 2025: 981m (+31.4% QoQ)
2025 was $NBIS ramp year:
Q4 2024: 37.9m
Q1: 50.9m (+34.3% QoQ)
Q2: 105.1m (+106.5% QoQ)
Q3: 146.1m (+39% QoQ)
Q4: 227.7m (+55.9% QoQ)
Q1 2026: 399m (+75.2% QoQ)
What Will Move the Stock
To look what will move Neocloud stocks, I admit that $NBIS has done a fantastic job this year. What $NBIS executed well objectively was:
1. Sign a 12B Meta Contract with +15B Extension Option
2. Back up their capability to fulfill the contract by hitting revenue numbers and critically showing acceleration in revenue growth. Observe how $NBIS Q1 2026 earnings show an acceleration to 75.2% revenue growth.
I monitor the whole industry to figure out what's going on and for $NBIS, I got to give credit where credit is due, $NBIS put up the GPUs and in this market it doesn't matter if you pay colocation or whatever, getting the GPUs up and showing revenue growth is what the market wants to see from early stage Neoclouds.
IREN's Revenue
IREN's ramp was suppose to start in Q4 2025 but really it's Q1 2026 because we couldn't get GPUs delivered on time due to HBM shortgage which snowballed the whole ramp process back.
Using currently delivery guidance, here are my calculations for the next few quarters revenue:
Q1: 33.6m
Q2: 100.8m (200% QoQ)
Q3: 207.6m (106% QoQ)
Q4: 385.4m (85.6% QoQ)
Q1: 843.8m (118.9% QoQ)
Q2: 1454.7m (72.4% QoQ)
Q1 Calculations: Reported in Q1 earnings
Q2 Calculations: PG exited Q1 with 307m run rate (page 19 of 10Q in source 1 - also screenshotted 1st picture) from the financial digging that @_Sgr_A_Star did and should be at 500m run rate by end of quarter. With linear ramp, (307+500/2)/4-qtr = 100m.
Q3 Calculations: 125m from PG and H1 Handoff stated by Dan to be in Q3 which I will take to be July. I'll take August + Sept of 124m from H1 so 124m * 2/3 = 82.6m.
Q4 Calculations: 125m from PG; 124m from H1; assuming we get Mackenzie handed half way through the quarter = 432.8/4qtr/2-halfway = 54.1m; assuming we get 2/3 duration of H2 and 1/3 duration of H3 and H4 get's handed over at the very end of the quarter, we only count the 2/3 H2+ 1/3 H3 = 124m. Sum = 427.1m
Q1 2027 Calculations: 125m from PG, 485m from H1-4, 108.2m from Mackenzie, CF = 40.6m, half duration of Nvidia Childress site = 85m.
Q2 2027 Calculations: 125m from PG, 485m from H1-4, 108.2m from Mackenzie, 40.6m from CF, 170m from Nvidia Childress, 1/2 duration of Block 7-9 is 330.9m, SW1 50MW IT is 195m
Share Price
IREN's ramp started late but having the power supply abundantly clear, means that the ramp can sustain high % growth for longer because power is not the bottleneck.
If IREN can have report 385.4m quarterly revenue for Q4 2026 it will mirror NBIS 399m quarter where most of its revenue was either H100/H200s and bare metal to MSFT with colocation payments so margin are similar. With an SW1 contract in hand, it would match the Meta 12B with potential 15B extension contract NBIS has. In this case, market is giving NBIS an 55B valuation.
Let's give 5B for NBIS 25% Clickhouse stake at 20B next round valuation even though at 15B valuation now. The rest of NBIS's subsidiary + the power they secured is rough equal to value of IREN's power portfolio (I know must IREN investors wouldn't make this trade off let's just call this even to make comparisons easy, you can do whatever adjustments you want). IREN Q4 2026 report will be which would be 50B (current NBIS market cap) / 22.89B (current IREN MCap) * 64.07 (current IREN stock price) = $139.95/share. Q4 earnings is early Feb 2027 but IREN also has higher sustain revenue growth rates due to it's power abundance but since Q4 earnings is Feb 2027, let's take 20% of for the time delta between EOY 2026 and earnings report for Q4 2026 to have a target of $111.96 share price for EOY 2026.
The really strong year for IREN will be 2027 as it sustains high growth rate and not be stuck in early ramp pains.
Sources
(1) iren.gcs-web.com/static-files/3…
Revenue
No doubt $IREN is rich in power, what $IREN investors need to focus on is revenue. The demand is certainly there, it's about getting GPUs online. Even the rate at which $IREN can sign contract is bottleneck by getting GPUs online because getting GPUs online derisk your ability to meet timelines to sign the next contract.
So what you see in $CRWV is that at the beginning they were really fast and then started slowing down. At first, a Neocloud will 2x revenue every quarter but then the ramp are bounded by the physical world. The benefit of having alot of power is that you'll be able to keep growing at a high rate until you ran out of power or your colocation provider(s) hits their limits. However, that's not $IREN's challenge. $IREN needs to get it's ramp -> cashflow -> ramp feedback loop going. Getting slow GPUs slowed down the whole ramp as you are doing theory based preparation for your datacenter until you get the GPUs and then problems can be uncovered sequentially.
Benchmarks
2024 was $CRWV's ramp year:
Q4 2023: 116m
Q1: 188.7m (+62.7% QoQ)
Q2: 395.4m (+109.5% QoQ)
Q3: 583.9m (+47.7% QoQ)
Q4: 747m (+27.9% QoQ)
Q1 2025: 981m (+31.4% QoQ)
2025 was $NBIS ramp year:
Q4 2024: 37.9m
Q1: 50.9m (+34.3% QoQ)
Q2: 105.1m (+106.5% QoQ)
Q3: 146.1m (+39% QoQ)
Q4: 227.7m (+55.9% QoQ)
Q1 2026: 399m (+75.2% QoQ)
What Will Move the Stock
To look what will move Neocloud stocks, I admit that $NBIS has done a fantastic job this year. What $NBIS executed well objectively was:
1. Sign a 12B Meta Contract with +15B Extension Option
2. Back up their capability to fulfill the contract by hitting revenue numbers and critically showing acceleration in revenue growth. Observe how $NBIS Q1 2026 earnings show an acceleration to 75.2% revenue growth.
I monitor the whole industry to figure out what's going on and for $NBIS, I got to give credit where credit is due, $NBIS put up the GPUs and in this market it doesn't matter if you pay colocation or whatever, getting the GPUs up and showing revenue growth is what the market wants to see from early stage Neoclouds.
IREN's Revenue
IREN's ramp was suppose to start in Q4 2025 but really it's Q1 2026 because we couldn't get GPUs delivered on time due to HBM shortgage which snowballed the whole ramp process back.
Using currently delivery guidance, here are my calculations for the next few quarters revenue:
Q1: 33.6m
Q2: 100.8m (200% QoQ)
Q3: 207.6m (106% QoQ)
Q4: 385.4m (85.6% QoQ)
Q1: 843.8m (118.9% QoQ)
Q2: 1454.7m (72.4% QoQ)
Q1 Calculations: Reported in Q1 earnings
Q2 Calculations: PG exited Q1 with 307m run rate (page 19 of 10Q in source 1 - also screenshotted 1st picture) from the financial digging that @_Sgr_A_Star did and should be at 500m run rate by end of quarter. With linear ramp, (307+500/2)/4-qtr = 100m.
Q3 Calculations: 125m from PG and H1 Handoff stated by Dan to be in Q3 which I will take to be July. I'll take August + Sept of 124m from H1 so 124m * 2/3 = 82.6m.
Q4 Calculations: 125m from PG; 124m from H1; assuming we get Mackenzie handed half way through the quarter = 432.8/4qtr/2-halfway = 54.1m; assuming we get 2/3 duration of H2 and 1/3 duration of H3 and H4 get's handed over at the very end of the quarter, we only count the 2/3 H2+ 1/3 H3 = 124m. Sum = 427.1m
Q1 2027 Calculations: 125m from PG, 485m from H1-4, 108.2m from Mackenzie, CF = 40.6m, half duration of Nvidia Childress site = 85m.
Q2 2027 Calculations: 125m from PG, 485m from H1-4, 108.2m from Mackenzie, 40.6m from CF, 170m from Nvidia Childress, 1/2 duration of Block 7-9 is 330.9m, SW1 50MW IT is 195m
Share Price
IREN's ramp started late but having the power supply abundantly clear, means that the ramp can sustain high % growth for longer because power is not the bottleneck.
If IREN can have report 385.4m quarterly revenue for Q4 2026 it will mirror NBIS 399m quarter where most of its revenue was either H100/H200s and bare metal to MSFT with colocation payments so margin are similar. With an SW1 contract in hand, it would match the Meta 12B with potential 15B extension contract NBIS has. In this case, market is giving NBIS an 55B valuation.
Let's give 5B for NBIS 25% Clickhouse stake at 20B next round valuation even though at 15B valuation now. The rest of NBIS's subsidiary + the power they secured is rough equal to value of IREN's power portfolio (I know must IREN investors wouldn't make this trade off let's just call this even to make comparisons easy, you can do whatever adjustments you want). IREN Q4 2026 report will be which would be 50B (current NBIS market cap) / 22.89B (current IREN MCap) * 64.07 (current IREN stock price) = $139.95/share. Q4 earnings is early Feb 2027 but IREN also has higher sustain revenue growth rates due to it's power abundance but since Q4 earnings is Feb 2027, let's take 20% of for the time delta between EOY 2026 and earnings report for Q4 2026 to have a target of $111.96 share price for EOY 2026.
The really strong year for IREN will be 2027 as it sustains high growth rate and not be stuck in early ramp pains.
Sources
(1) iren.gcs-web.com/static-files/3…
6
@GavinSBaker
Jim Liu@jiahanjimliu
$IREN: Rapidly Rising Cost of BTM Gas Turbine
Gas turbines and associated pipeline and equipment are now 3m-4m/MW based off Wells Farm research from @ShanuMathew93! That's not even including the cost of natural gas. Without a major pipeline or source of natural gas nearby, liquified natural gas is needed which is 15 cents/kWH.
The market has rewarded $BE for producing MWs in the form of fuel cells which are good for 5 years and BE sells them in 10 year service contracts aka fuel cells as a service. Eventually IREN's MWs will be valued too. I mean why pay 4B/GW in BTM Gas Turbine equipment when you can get grid connected power at rates opex same or lower than generating from natural gas in many states?
Margin calculation wise, grid connected datacenter build is 10m/MW so BTM Gas Turbines is an increase of 30-40% on upfront capex which eats directly into profits.
Heavy Redundancy Factor
With a redundancy factor of 30-40% needed, many MWs of capacity can't even be used. Who cares if you have a 1.1 PUE datacenter if you are BTM Gas Turbines? The 1.1 PUE datacenter effectively became 1.1 * 1.4 = 1.54 PUE. And yes it's multiplicative because you need redundancy in the gas turbines for the AUX load like cooling too.
People will figure out this is much more important than the cost of electricity. BTM is a very expensive solution.
I never believed in the tokens/MW metric because really it should be tokens/(GPU+HBM) because cost of GPUs and HBM which is still the dominant factor here but BTM screws over any token/MW metric too once you factor in redundancy needed.
Why were Gas Turbines considered before $BE Fuel Cells and Smaller Engines like Reciprocal Engines?
Gas power plants last 25 years. In the case of NBIS's 2.6B for 328MW, BE Fuel Cells are 7.4m/MW for 10 years only. When normalized to 20 lifetime of datacenter, gas turbines are 2.8m/MW-20yrs while BE Fuel Cells are 14.4m/MW-20yrs!
Fuel Cells also have a redundancy factor 31% as NBIS is getting 250MW of usable power and 78MW of the fuel cells are for redundancy. That makes a 1.1 PUE datacenter effectively 1.1 * 1.31 = 1.44!
Smaller engines are in-between cost curve of Gas Turbines and Fuel Cells which is why NBIS attempted to use Bergen Cruise Ship Engines before they switch to $BE Fuel Cells.
Impact on IREN
Market doesn't reward IREN for hoarding GWs. When IREN converts the GWs to compute and the cost effectiveness of grid connected power over BTM Gas Turbines or Fuel Cells is apparent, IREN will be rewarded. Jane Street Algos and Wall Street Analyst rewards hard metrics like revenue, margins, and profit, not capacity to be built in the future like MWs.
IREN is seeing painful delays now on GPU deliveries which back up their entire datacenter bring up schedule. On the backend, however, having this huge power portfolio will allow IREN to build out the fastest unburdened by BTM buildout time on top of DC buildout time.
Thanks to @alanbialo for pointing out this post to me.
Gas turbines and associated pipeline and equipment are now 3m-4m/MW based off Wells Farm research from @ShanuMathew93! That's not even including the cost of natural gas. Without a major pipeline or source of natural gas nearby, liquified natural gas is needed which is 15 cents/kWH.
The market has rewarded $BE for producing MWs in the form of fuel cells which are good for 5 years and BE sells them in 10 year service contracts aka fuel cells as a service. Eventually IREN's MWs will be valued too. I mean why pay 4B/GW in BTM Gas Turbine equipment when you can get grid connected power at rates opex same or lower than generating from natural gas in many states?
Margin calculation wise, grid connected datacenter build is 10m/MW so BTM Gas Turbines is an increase of 30-40% on upfront capex which eats directly into profits.
Heavy Redundancy Factor
With a redundancy factor of 30-40% needed, many MWs of capacity can't even be used. Who cares if you have a 1.1 PUE datacenter if you are BTM Gas Turbines? The 1.1 PUE datacenter effectively became 1.1 * 1.4 = 1.54 PUE. And yes it's multiplicative because you need redundancy in the gas turbines for the AUX load like cooling too.
People will figure out this is much more important than the cost of electricity. BTM is a very expensive solution.
I never believed in the tokens/MW metric because really it should be tokens/(GPU+HBM) because cost of GPUs and HBM which is still the dominant factor here but BTM screws over any token/MW metric too once you factor in redundancy needed.
Why were Gas Turbines considered before $BE Fuel Cells and Smaller Engines like Reciprocal Engines?
Gas power plants last 25 years. In the case of NBIS's 2.6B for 328MW, BE Fuel Cells are 7.4m/MW for 10 years only. When normalized to 20 lifetime of datacenter, gas turbines are 2.8m/MW-20yrs while BE Fuel Cells are 14.4m/MW-20yrs!
Fuel Cells also have a redundancy factor 31% as NBIS is getting 250MW of usable power and 78MW of the fuel cells are for redundancy. That makes a 1.1 PUE datacenter effectively 1.1 * 1.31 = 1.44!
Smaller engines are in-between cost curve of Gas Turbines and Fuel Cells which is why NBIS attempted to use Bergen Cruise Ship Engines before they switch to $BE Fuel Cells.
Impact on IREN
Market doesn't reward IREN for hoarding GWs. When IREN converts the GWs to compute and the cost effectiveness of grid connected power over BTM Gas Turbines or Fuel Cells is apparent, IREN will be rewarded. Jane Street Algos and Wall Street Analyst rewards hard metrics like revenue, margins, and profit, not capacity to be built in the future like MWs.
IREN is seeing painful delays now on GPU deliveries which back up their entire datacenter bring up schedule. On the backend, however, having this huge power portfolio will allow IREN to build out the fastest unburdened by BTM buildout time on top of DC buildout time.
Thanks to @alanbialo for pointing out this post to me.
7
@GavinSBaker
Jim Liu@jiahanjimliu
$IREN: Road to Inference SaaS
Prior to Token Factory, $NBIS was partner with Hugging Face for Open Source Models. With Mirantis acquisition closing in August, $IREN has the following path to be a inference provider.
1. Integrate Mirantis k0rdent + Kubernetes with KServe + vLLM. Note that k0rdent is reponsible for scalable inference hosting.
2. Partner with Hugging Face for open source models with quantized versions when appropriate, optimum optimized models, optimizations for continuous batching, paged attention, tensor parallelism.
3. Add billing and enterprise security.
4. Optional: can acquire startups similar to how Nebius acquired Eigen AI / Clarifai for further optimizations. Note that Eigen AI / Clarifai have very high benchmarks because they are optimized for being acquired and achieved those benchmarks with minimal batch size and uses more GPU slices to achieve their benchmarks. Fireworks/Coreweave have to deal with unit economics. In practice when unit economics matter, Fireworks has the best inference capabilities among Neoclouds. CRWV and Nebius with their acquisitions have strong inference offerings. HS also have open source inference offerings but only FireworksAI inference is also sold on HS platforms.
Nebius launched Token Factory in Nov 2025, so IREN would be about 1-1.5 years behind after closing Mirantis acquisition in August 2026 and integrating k0rdent with KServe + vLLM.
Prior to Token Factory, $NBIS was partner with Hugging Face for Open Source Models. With Mirantis acquisition closing in August, $IREN has the following path to be a inference provider.
1. Integrate Mirantis k0rdent + Kubernetes with KServe + vLLM. Note that k0rdent is reponsible for scalable inference hosting.
2. Partner with Hugging Face for open source models with quantized versions when appropriate, optimum optimized models, optimizations for continuous batching, paged attention, tensor parallelism.
3. Add billing and enterprise security.
4. Optional: can acquire startups similar to how Nebius acquired Eigen AI / Clarifai for further optimizations. Note that Eigen AI / Clarifai have very high benchmarks because they are optimized for being acquired and achieved those benchmarks with minimal batch size and uses more GPU slices to achieve their benchmarks. Fireworks/Coreweave have to deal with unit economics. In practice when unit economics matter, Fireworks has the best inference capabilities among Neoclouds. CRWV and Nebius with their acquisitions have strong inference offerings. HS also have open source inference offerings but only FireworksAI inference is also sold on HS platforms.
Nebius launched Token Factory in Nov 2025, so IREN would be about 1-1.5 years behind after closing Mirantis acquisition in August 2026 and integrating k0rdent with KServe + vLLM.
8
@GavinSBaker
Jim Liu@jiahanjimliu
Space DC: Napkin Math Debunked
Preface
To preface, I'm a Elon fan, had my earliest big stock win via all-in $TSLA and used to work at Tesla's Palo Alto office. I do think Space DCs will be a part of the solution to alleviate the power bottleneck but I'll show that they are not cheaper than terrestrial DCs.
To demonstrate, I'll only need to add one reason to @WR4NYGov's (Warren Redlich) napkin math to debunk and I won't need any. Warren is one of $TSLA investors community's early core members much like @FransBakker9812 is for $IREN.
I don't doubt Elon's engineering capabilities, so I'm going to take all of Warren's assumptions on radiant cooling and radiactive shielding working, reusable Falcon heavy to drive down launch costs, etc. Provided that all the engineering and cost optimizations works, Warren comes out with $10m to power on a Vera Rubin rack in space vs $12m per rack on the ground.
Key Point
The key point Warren misses is that terrestrial DC can be retrofitted to fit the new generation at 3-4m / rack while Space DCs need a relaunch, however reusable the rocket is. Using Warren's numbers, we will show the 20 year cost of Space DCs and terrestrial DCs under both cases of 5 year GPU lifetime and 7 year GPU lifetime.
For 5 year GPU lifetime, Space DC will have 4 launches over 20 years = 40m. For terrestrial DCs, it will be 12m initial build cost and (3.5m * 3 retrofits) = 22.5m.
Now, longer GPU lifetimes help the case for Space DCs. For 7 year GPU lifetime, Space DC will have 4 launches over 21 years = 30m. For terrestrial DCs, it will be 12m initial build cost and (3.5m * 2) retrofits) = 19m.
Conclusion
You see, I am giving Space DC maximal benefit of the doubt that all the engineering works and working off Warren's assumption that the initial cost including solar, radiant cooling of Space DCs is cheaper than terrestrial DCs to begin with.
Sure, Warren has stated that he hasn't counted the savings in electricity cost yet, but for companies like $IREN with cheap re-usable power in West Texas, South Australia and relatively smaller sites in Spain and British Columbia, electricity costs aren't the difference maker of 17.5m per rack (5 year GPU) or 11m per rack (7 year GPU).
So yes Space DCs are cost competitive with novel terrestrial solutions like $BE Fuel Cells but not with grid connected power. And none of this argument needs to touch on cost of redundancy needed in space since you can't economically make repairs in space, cost of insuring expensive Vera Rubins in space, and not to mention maintenance. Maintenance is actually a big one, Starlink needs little maintenance since it runs the same loop of software while GPU Cloud runs a diverse workload that many need all types of manual intervention. My wife works at Oracle Cloud Infra and it's not all to uncommon to call the Datacenter technicians.
Preface
To preface, I'm a Elon fan, had my earliest big stock win via all-in $TSLA and used to work at Tesla's Palo Alto office. I do think Space DCs will be a part of the solution to alleviate the power bottleneck but I'll show that they are not cheaper than terrestrial DCs.
To demonstrate, I'll only need to add one reason to @WR4NYGov's (Warren Redlich) napkin math to debunk and I won't need any. Warren is one of $TSLA investors community's early core members much like @FransBakker9812 is for $IREN.
I don't doubt Elon's engineering capabilities, so I'm going to take all of Warren's assumptions on radiant cooling and radiactive shielding working, reusable Falcon heavy to drive down launch costs, etc. Provided that all the engineering and cost optimizations works, Warren comes out with $10m to power on a Vera Rubin rack in space vs $12m per rack on the ground.
Key Point
The key point Warren misses is that terrestrial DC can be retrofitted to fit the new generation at 3-4m / rack while Space DCs need a relaunch, however reusable the rocket is. Using Warren's numbers, we will show the 20 year cost of Space DCs and terrestrial DCs under both cases of 5 year GPU lifetime and 7 year GPU lifetime.
For 5 year GPU lifetime, Space DC will have 4 launches over 20 years = 40m. For terrestrial DCs, it will be 12m initial build cost and (3.5m * 3 retrofits) = 22.5m.
Now, longer GPU lifetimes help the case for Space DCs. For 7 year GPU lifetime, Space DC will have 4 launches over 21 years = 30m. For terrestrial DCs, it will be 12m initial build cost and (3.5m * 2) retrofits) = 19m.
Conclusion
You see, I am giving Space DC maximal benefit of the doubt that all the engineering works and working off Warren's assumption that the initial cost including solar, radiant cooling of Space DCs is cheaper than terrestrial DCs to begin with.
Sure, Warren has stated that he hasn't counted the savings in electricity cost yet, but for companies like $IREN with cheap re-usable power in West Texas, South Australia and relatively smaller sites in Spain and British Columbia, electricity costs aren't the difference maker of 17.5m per rack (5 year GPU) or 11m per rack (7 year GPU).
So yes Space DCs are cost competitive with novel terrestrial solutions like $BE Fuel Cells but not with grid connected power. And none of this argument needs to touch on cost of redundancy needed in space since you can't economically make repairs in space, cost of insuring expensive Vera Rubins in space, and not to mention maintenance. Maintenance is actually a big one, Starlink needs little maintenance since it runs the same loop of software while GPU Cloud runs a diverse workload that many need all types of manual intervention. My wife works at Oracle Cloud Infra and it's not all to uncommon to call the Datacenter technicians.
9
@GavinSBaker $IREN benefits more from Frontier Labs than Open Source currently. Frontier Labs have model-hardware co-design advantage over Open Source and this will only become more significant.
Jim Liu@jiahanjimliu
Frontier Labs vs Open Source
@GavinSBaker explains a critical advantage frontier labs have over open source: model-hardware co-design is becoming more important (FinX knows the term vertical integration).
Co-design means you design the model for hardware and design the hardware for the model. Chinese open source labs cannot be designing their model architecture for American silicon. American silicon will not design their silicon for Chinese open source. Every top open source model is Chinese.
Model-hardware co-design is becoming critical in performance and efficiency. Open source and smaller companies cannot design their own wafers on the leading node nor can they impact major ASIC roadmaps. Meanwhile Anthropic has can directly request changes in Trainium’s design. Nvidia co-designs to OpenAI and Anthropic to make sure their models are performant on Nvidia GPUs.
What this means is that Anthropic and OpenAI will have higher intelligence per dollar-watt than open source. This will allow them to draw margins from efficiency and have total cost to run competitive to open source at the same time. Open source is not free, it needs expensive Nvidia hardware.
@GavinSBaker explains a critical advantage frontier labs have over open source: model-hardware co-design is becoming more important (FinX knows the term vertical integration).
Co-design means you design the model for hardware and design the hardware for the model. Chinese open source labs cannot be designing their model architecture for American silicon. American silicon will not design their silicon for Chinese open source. Every top open source model is Chinese.
Model-hardware co-design is becoming critical in performance and efficiency. Open source and smaller companies cannot design their own wafers on the leading node nor can they impact major ASIC roadmaps. Meanwhile Anthropic has can directly request changes in Trainium’s design. Nvidia co-designs to OpenAI and Anthropic to make sure their models are performant on Nvidia GPUs.
What this means is that Anthropic and OpenAI will have higher intelligence per dollar-watt than open source. This will allow them to draw margins from efficiency and have total cost to run competitive to open source at the same time. Open source is not free, it needs expensive Nvidia hardware.
10
@GavinSBaker
Jim Liu@jiahanjimliu
The bottleneck $IREN needs to overcome.
Having higher priority on GPUs via Nvidia Partnership will be critical for IREN. Unfortunately, 2026 GPUs are all allocated.
You can see here that used A100/H100/H200s are sold out with a backlog.
$IREN has secured GPUs for 3.7B ARR exiting 2026 but they are all coming later in each quarter.
Having higher priority on GPUs via Nvidia Partnership will be critical for IREN. Unfortunately, 2026 GPUs are all allocated.
You can see here that used A100/H100/H200s are sold out with a backlog.
$IREN has secured GPUs for 3.7B ARR exiting 2026 but they are all coming later in each quarter.
11
@GavinSBaker Dell Software Capabilities:
Jim Liu@jiahanjimliu
Question for @FransBakker9812, @Agrippa_Inv - is $IREN using Dell Software for some customers?
Ben is very credible and although he has $IREN rated lower than $CRWV and $NBIS by a lot, Dell is rated higher than all 3.
Software fees for Dell is most likely much cheaper than Colocation fees, and favorable if software is superior.
Overall, I think $IREN’s bread and butter customer should be FireworksAI and Anthropic that bring their own software stack
Ben is very credible and although he has $IREN rated lower than $CRWV and $NBIS by a lot, Dell is rated higher than all 3.
Software fees for Dell is most likely much cheaper than Colocation fees, and favorable if software is superior.
Overall, I think $IREN’s bread and butter customer should be FireworksAI and Anthropic that bring their own software stack
12
@GavinSBaker Databricks CEO Interview on AI Impact on SaaS:
Jim Liu@jiahanjimliu
Neoclouds: AI Impact on SaaS
Databricks is a SaaS itself but is adapting well to AI. Databricks CEO, Ali Ghodsi, is probably the one of the best expert to ask the question of AI impact on SaaS.
Although we haven’t seen anywhere near the full effects yet, it’s important to keep what Ali Ghodsi says:
“Writing software is ten times faster now... your margin is my opportunity”.
Databricks is a SaaS itself but is adapting well to AI. Databricks CEO, Ali Ghodsi, is probably the one of the best expert to ask the question of AI impact on SaaS.
Although we haven’t seen anywhere near the full effects yet, it’s important to keep what Ali Ghodsi says:
“Writing software is ten times faster now... your margin is my opportunity”.
13
@GavinSBaker
Jim Liu@jiahanjimliu
$IREN: Horizon 1-4 Tech Deep Dive and Software Layers Explained
IREN doesn't directly reveal technical specs about the Horizon 1-4 datacenters but from the specs and Frans pictures we can deduce that they are using Nvidia Spectum-XGS architecture.
1. Each Horizon is 3 buildings. 2 Horizons or 6 buildings make up a supercluster (Frans pictures).
2. Networking core as a scale-across network layer across buildings (Frans pictures).
3. Microsoft will use these for training (picture 1, also source 3).
Spectrum-XGS allows GPU networking to scale beyond one building and form synchronized clusters across multiple buildings using Nvidia system software:
1. Distance-aware congestion control via DOCA SPC-X CC (1).
2. Topology-aware routing via LAG hash randomizer for adaptive routing.
3. Firmware support from Spectrum-X RA 2.1 in Spectrum-X Ethernet switches and ConnectX-8 SuperNICs.
This involves integrating Nvidia's Kubernetes NIC Configuration. IREN's had Kubernetes and infrastructure SDE's since well before Mirantis:
Justing Nearing: linkedin.com/in/justinneari…
Tyrell Kumlin: linkedin.com/in/tyrell-kuml…
Ivan Temchenko: linkedin.com/in/itemchenko/z
Igor Widlinski: linkedin.com/in/igorwidlins…
Peter Dou: linkedin.com/in/igorwidlins…
Even though IREN has Kubernetes setup capabilities prior, Mirantis will be a good acquisition for orchestration. Endgame I think IREN should consider Baseten or Modal which will put make them competitive with NBIS. In AI Software Platforms, I would rank Fireworks == Databricks > CRWV > NBIS == Baseten > Modal == TogetherAI > Lambda. Fireworks turned down a CRWV acquisition back in the day and has only grown stronger.
IREN's Horizon 1-4 datacenters are state of the art but not singular. CRWV will also be using Nvidia Spectrum-XGS scale across design (4). Spectrum-XGS is significant because it doubles (1.9x) the performance of NCCL (4), one of Nvidia's most fundamental libraries, as important as CUDA.
IREN vs Mirantis vs Baseten/Modal Software:
1. IREN's previous software capabilities for bare metal is to configure Kubernetes to be able to use the networking technologies discussed above as well as other technologies. Operationally, this involves updating firmware, updating non-volatile memory configuraitons, tune RoCE and SPC-X congestion-control settings to finally expose NIC virtual function to the Kubernetes pods.
2. Whereas IREN configures Kubernetes pod as described above, what Mirantis brings is Kubernetes orchestration which manages workload placement and resource allocation. This is a key foundation for auto inference scaling.
3. What Baseten/Modal brings is serverless or auto-scaling inference, optimization of GPU utilization, observability, and security. This means that developers do have to work with Kubernetes or any infrastructure and can deal with models, APIs.
Baseten is the stronger of the two candidates I like but quite more expensive. Baseten raised a 13B valuation and Modal is at a 4.65B valuation.
Fireworks, Databricks, Baseten, Modal, TogetherAI are asset-super-lite meaning they don't own even the GPUs and rent bare metal GPUs from hyperscalers and sell managed AI services on their AI platforms. Fireworks/TogetherAI also buy IREN bare metal.
AI platforms moves fast, FireworksAI has recently released managed training which in finX terms, is like Token Factory but for training custom models from open source base (5). I expect CRWV, NBIS and other Neoclouds to follow suit.
IREN doesn't directly reveal technical specs about the Horizon 1-4 datacenters but from the specs and Frans pictures we can deduce that they are using Nvidia Spectum-XGS architecture.
1. Each Horizon is 3 buildings. 2 Horizons or 6 buildings make up a supercluster (Frans pictures).
2. Networking core as a scale-across network layer across buildings (Frans pictures).
3. Microsoft will use these for training (picture 1, also source 3).
Spectrum-XGS allows GPU networking to scale beyond one building and form synchronized clusters across multiple buildings using Nvidia system software:
1. Distance-aware congestion control via DOCA SPC-X CC (1).
2. Topology-aware routing via LAG hash randomizer for adaptive routing.
3. Firmware support from Spectrum-X RA 2.1 in Spectrum-X Ethernet switches and ConnectX-8 SuperNICs.
This involves integrating Nvidia's Kubernetes NIC Configuration. IREN's had Kubernetes and infrastructure SDE's since well before Mirantis:
Justing Nearing: linkedin.com/in/justinneari…
Tyrell Kumlin: linkedin.com/in/tyrell-kuml…
Ivan Temchenko: linkedin.com/in/itemchenko/z
Igor Widlinski: linkedin.com/in/igorwidlins…
Peter Dou: linkedin.com/in/igorwidlins…
Even though IREN has Kubernetes setup capabilities prior, Mirantis will be a good acquisition for orchestration. Endgame I think IREN should consider Baseten or Modal which will put make them competitive with NBIS. In AI Software Platforms, I would rank Fireworks == Databricks > CRWV > NBIS == Baseten > Modal == TogetherAI > Lambda. Fireworks turned down a CRWV acquisition back in the day and has only grown stronger.
IREN's Horizon 1-4 datacenters are state of the art but not singular. CRWV will also be using Nvidia Spectrum-XGS scale across design (4). Spectrum-XGS is significant because it doubles (1.9x) the performance of NCCL (4), one of Nvidia's most fundamental libraries, as important as CUDA.
IREN vs Mirantis vs Baseten/Modal Software:
1. IREN's previous software capabilities for bare metal is to configure Kubernetes to be able to use the networking technologies discussed above as well as other technologies. Operationally, this involves updating firmware, updating non-volatile memory configuraitons, tune RoCE and SPC-X congestion-control settings to finally expose NIC virtual function to the Kubernetes pods.
2. Whereas IREN configures Kubernetes pod as described above, what Mirantis brings is Kubernetes orchestration which manages workload placement and resource allocation. This is a key foundation for auto inference scaling.
3. What Baseten/Modal brings is serverless or auto-scaling inference, optimization of GPU utilization, observability, and security. This means that developers do have to work with Kubernetes or any infrastructure and can deal with models, APIs.
Baseten is the stronger of the two candidates I like but quite more expensive. Baseten raised a 13B valuation and Modal is at a 4.65B valuation.
Fireworks, Databricks, Baseten, Modal, TogetherAI are asset-super-lite meaning they don't own even the GPUs and rent bare metal GPUs from hyperscalers and sell managed AI services on their AI platforms. Fireworks/TogetherAI also buy IREN bare metal.
AI platforms moves fast, FireworksAI has recently released managed training which in finX terms, is like Token Factory but for training custom models from open source base (5). I expect CRWV, NBIS and other Neoclouds to follow suit.
14
@GavinSBaker
Jim Liu@jiahanjimliu
Why $IREN is Pivoting Half Their Capacity to Software Cloud instead of Enterprise Bare Metal
Thanks to @alanbialo for pointing out this post. Nvidia has alot of say in Neoclouds' roadmaps. Some behind the scene politics from @SemiAnalysis_ shed light on what's happening:
1. "Many neocloud executives we spoke with feel that if they have non-NVIDIA networking gear on their cluster, or if their cloud has an AMD GPU or TPU offering, NVIDIA will retaliate."
2. "They feel that retaliation includes not giving early allocation or no longer supporting a potential IPO/VC raise."
3. "for neoclouds, executives feel NVIDIA has been using high-pressure tactics to keep them NVIDIA-only."
This behavior is not new in the tech industry as Apple is notorious for bullying their suppliers (1), sometimes into negative margins. Intel, IBM also did so when they were dominant. Jensen is, overall, a more generous and long-term outlooking partner than Apple, Intel or IBM. Jensen takes but Jensen will also giveth. Jensen has to stack the industry in his mold but will allow his players to profit as indiciated by their long standing partnership with TSMC.
The sober reality is that for Neoclouds it's important to pledge your allegiance to Nvidia to get your GPUs. By building out a AI Software Platform and pivotting half it's capacity to small enterprises, $IREN is serving as an alternative store front for Nvidia. $IREN will be getting priority GPU allocations as it aligns with Nvidia incentives.
Jensen Giveth
Pointed out by @HowAreYou818, Firmus AI Cloud is forming a partnership to "to procure NVIDIA infrastructure for AI-Native, enterprise and ISV customers, through economic alignment with a revenue-sharing and credit-support model" (2). This is for 170k GPUs on 360MW (2) with "Firmus expects to receive between US$25 and US$30 billion from committed offtake agreements during the first six years of the partnership."
One option is that $IREN could look to partner for a similar agreement. It's not clear what 25B-30B from offtake agreement entails as it doesn't make sense for straight revenue or profit. Most liley some JV with GPU lease from Nvidia. With Nvidia involvement, building out IREN's 5.8GW becomes more likely and would convert to a 400B-480B offtake agreement over 6 years.
Thanks to @alanbialo for pointing out this post. Nvidia has alot of say in Neoclouds' roadmaps. Some behind the scene politics from @SemiAnalysis_ shed light on what's happening:
1. "Many neocloud executives we spoke with feel that if they have non-NVIDIA networking gear on their cluster, or if their cloud has an AMD GPU or TPU offering, NVIDIA will retaliate."
2. "They feel that retaliation includes not giving early allocation or no longer supporting a potential IPO/VC raise."
3. "for neoclouds, executives feel NVIDIA has been using high-pressure tactics to keep them NVIDIA-only."
This behavior is not new in the tech industry as Apple is notorious for bullying their suppliers (1), sometimes into negative margins. Intel, IBM also did so when they were dominant. Jensen is, overall, a more generous and long-term outlooking partner than Apple, Intel or IBM. Jensen takes but Jensen will also giveth. Jensen has to stack the industry in his mold but will allow his players to profit as indiciated by their long standing partnership with TSMC.
The sober reality is that for Neoclouds it's important to pledge your allegiance to Nvidia to get your GPUs. By building out a AI Software Platform and pivotting half it's capacity to small enterprises, $IREN is serving as an alternative store front for Nvidia. $IREN will be getting priority GPU allocations as it aligns with Nvidia incentives.
Jensen Giveth
Pointed out by @HowAreYou818, Firmus AI Cloud is forming a partnership to "to procure NVIDIA infrastructure for AI-Native, enterprise and ISV customers, through economic alignment with a revenue-sharing and credit-support model" (2). This is for 170k GPUs on 360MW (2) with "Firmus expects to receive between US$25 and US$30 billion from committed offtake agreements during the first six years of the partnership."
One option is that $IREN could look to partner for a similar agreement. It's not clear what 25B-30B from offtake agreement entails as it doesn't make sense for straight revenue or profit. Most liley some JV with GPU lease from Nvidia. With Nvidia involvement, building out IREN's 5.8GW becomes more likely and would convert to a 400B-480B offtake agreement over 6 years.
15
@GavinSBaker
Jim Liu@jiahanjimliu
Neoclouds: Exemplar Cloud Status
I never took Exemplar Cloud Status seriously and had no doubt that $IREN could easily catch up to get the Exemplar Cloud Status badge.
Nvidia providers the configuration agents like for Kubernetes Networking Configuration Operator here: github.com/Mellanox/nic-c…
Nvidia is doing all the heavy lifting, all Neoclouds like $CRWV, $NBIS, $IREN needs to do is hire a strong team of DevOps engineers:
Justin Nearing: linkedin.com/in/justinneari…
Tyrell Kumlin: linkedin.com/in/tyrell-kuml…
Ivan Temchenko: linkedin.com/in/itemchenko/…
Igor Widlinski: linkedin.com/in/igorwidlins…
Peter Dou: linkedin.com/in/ziqi-dou/
Nvidia is doing the rocket science. Neoclouds do mostly integration outside of craft in model inference optimization. Don't let the software narrative scare you. $IREN shows how well Nvidia is doing leveling the playing field as it achieves top performance benchmarks on B300s to achieve Nvidia Exemplar Cloud Status.
I never took Exemplar Cloud Status seriously and had no doubt that $IREN could easily catch up to get the Exemplar Cloud Status badge.
Nvidia providers the configuration agents like for Kubernetes Networking Configuration Operator here: github.com/Mellanox/nic-c…
Nvidia is doing all the heavy lifting, all Neoclouds like $CRWV, $NBIS, $IREN needs to do is hire a strong team of DevOps engineers:
Justin Nearing: linkedin.com/in/justinneari…
Tyrell Kumlin: linkedin.com/in/tyrell-kuml…
Ivan Temchenko: linkedin.com/in/itemchenko/…
Igor Widlinski: linkedin.com/in/igorwidlins…
Peter Dou: linkedin.com/in/ziqi-dou/
Nvidia is doing the rocket science. Neoclouds do mostly integration outside of craft in model inference optimization. Don't let the software narrative scare you. $IREN shows how well Nvidia is doing leveling the playing field as it achieves top performance benchmarks on B300s to achieve Nvidia Exemplar Cloud Status.
16
@GavinSBaker Industry Competitor: $NBIS
Jim Liu@jiahanjimliu
$NBIS: Multiple people have asked me recently about buying $NBIS at $200+. They have, obviously, read the multiple bull cases, but from me, they are looking to get an alternative point of view for balance. I'll layout the risks in a public post for transparency.
In all fairness, if you are reading this for the first time and want the bull case, these accounts have done a great job laying out the $NBIS bull case:
1. @daniel_koss
2. @itsalasdairmann
3. @MB_Hogan
4. @babyfolio
5. @EndicottInvests
I don't own any $NBIS but some $IREN whales do own a fair amount of $NBIS and have done well with $IREN in 2025 and $NBIS in 2026.
There's no doubt that in H1 2026, $NBIS has had the best stock performance among publicly traded Neoclouds, with concrete deliverables in 399m Q1 revenue on 82% QoQ growth (1) and 40% downpayment from MSFT (2). At 11-15m/MW-yr, 399m quarterly revenue equates to 106MW-145MW of build out depending on blend GPUs and IaaS-PaaS-SaaS. This buildout is still early and not bottlenecked by power. However, hot stock performance for one time frame doesn't determine the whole trajectory; an example is $SMCI vs $DELL and $HPE. What will de-risk $NBIS for me is completion of their Vineland, NJ datacenter in 2026 or early 2027.
I will go over two main topics: Power and Software.
1. Power
Nebius has more than 4GWs of secured power with a strategy of diversifying between colocation and self-owned over many sites to manage risk. Nebius refers to secure and contracted power as either PPA from the energy producer or interconnection agreement (IA) which is final approval needed for grid buildout. Nebius' BTM sites like Vineland, NJ do not need IA and also bypass the grid buildout time.
Not all Nebius sites are equal so I'll go through starting from the ones that are the best in terms of time to power without delays.
Mäntsälä
Nebius's Mäntsälä site is 75MW of active power (11). This datacenter from Nebius's Yandex days is grid connected, greenfield, and self-owned (11). The Finland datacenter boast a PUE of 1.13 (11) comparable with IREN's Prince George datacenter at 1.1 PUE (12), both taking advantage of the climate.
Iceland
10MW of colocation lock and loaded (26).
Kansas City, Missouri
5GW of Colocation from Patmos lock and loaded, potential expansion to 40MW (29).
Israel
Israel my favorite $NBIS site in terms of guaranteed path to power with no delays. Nebius signed a datacenter lease for 80MW in Modi'in in the Masmiyya and Beit Shemesh districts with expansion to 286MW (23). Some of the clusters will be used for Israel's national super computer (24) and Beit Shemesh is an important military district, previously target of Iran's military strikes but now de-risked as the Iran Wwar has ended (25).
Independence
Of Nebius's large sites in North America, I like the potential of Independence the best. It cleverly uses reopening an old power plant built in 1958 and closed in 2020 near by to provide power. The natural gas-fueled Blue Valley Power Plant is very close by Nebius' designated site so grid buildout requires less high voltage components.
This is similar to Elon's colossus site which uses a mix of reopened power plant and importing gas turbines from deconstruction power plants from abroad (22).
UK
Ark Data Centres colocation, 16MW agreed with expansion plans to 52MW (30).
Vineland
Nebius current active buildout is a Colcoation site in Vineland, NJ with "Nebius, becoming a DataOne tenant for the next 10 years" (7). From Nebius' 20-F at the end of 2024 (9) and F-3 mid 2025 (10), "The New Jersey site is a phased development scalable up to 300 MW, with initial capacity expected to be available in the second half of 2025." However, we are entering into the second half of 2026 and Vineland has just switched out their power supply from Bergen cruiseship engines with Skylea carbon capture (13) to $BE fuel cells (14).
Nebius is paying 2.6B for 328MW of $BE Fuel Cells as a Service for 10 years with 250MW of guaranteed capacity and 78MW of installed system capacity (14). Like all BTM power sources, redundancy factor is needed which is the 78MW and only 250MW goes to the GPU and Aux Loads. This is 10.4m/MW for 10 years or 1.04m/MW-yr and doesn't even include the natural gas (17). To put in perspective, the entire datacenter cost buildout 15m/MW for 20 years with 3-4m retorfit cost if a site has grid connected power.
There is a concern whether or not Nebius will get the $BE fuel cells this year as they are sold out for 2026 (16). This is not a concern and the fuel cells themselves will be delivered this year because Oracle encountered a delay for their NM site and $BE is rerouting the fuel cells from Oracle to Nebius (15).
As stated in the 6-K, Nebius may provide "alternative capacity" (31) for it's Microsoft contract while Vineland is under construction. This will work for several tranches but will not work forever as Vineland is 3.48B ARR and almost half of $NBIS 7-9B 2026 ARR guidance.
Lappeenranta
Nebius has labled Lappeenranta, Finland as a 310MW greenfield site. Lappeenaranta is a greenfield site belonging to Polarnode and Nebius is renting it as Colocation (18). @TheUglyBuckling has full details on this site: x.com/TheUglyBucklin…
Bethune, France
@InvestNorthwise does a good job of covering this one here: northwiseproject.com/nbis-bethune-f….
Birmingham, Alabama
300MW of brownfield development that got permitted right before the city passed it's datacenter moratorium. This site has broken ground and has a decent chance to completion as a BTM site despite lawsuit from two homeowners who are within 1000ft of the datacenter (27).
The major blocker from Birmingham succeeding as a grid connected site is that the permit to build a substation and switching station was denied and final (28). The workaround to this is obviously just like Vineland $BE fuel cells. This is why I see Vineland, NJ as a pivotal site for Nebius as alot of their sites' contingency plan is $BE fuel cells so Vineland is the litmus test for their ability to execute on $BE fuel cell BTM sites.
Pennsylvania
Nebius most recent site announcement is their 1.2GW Pennsylvania site which they revealed to Analyst is Pennsylvania is grid connected (3). By end of 2027, Pennsylvania will have 250MW-350MW with each year adding 300MW (5). From cross referencing total capacity, power ramp, and acreage of site, Nebius's Pennsylvania site is most likely Lower Mount Bethel Technology Center” in Northampton County (6). If so, Nebius has secured the power purchase agreement and still needs to get the interconnection agreement which can be delayed based on time to build transmission, substations, and grid. Regulations and local anti-datacenter activism is dealt with in parallel to optimize time to power.
Good news is PJM is reforming it's interconnection queue to be first-come first-serve (19). Bad news is despite that, PJM is one of the most congested buildouts from a supply chain perspective and "PJM Interconnection data – but the biggest delays are no longer occurring within the interconnection queue" (20).
"The bottleneck has shifted downstream. Transmission buildouts, substation capacity, and strained supply chains are now the primary obstacles to energizing projects. According to PJM, projects spent an average of more than three years reaching an interconnection service agreement, and another four years waiting to come online after approval" (20).
2. Software
A large part of Nebius market cap is attributed software. Although Nebius is way ahead of IREN in software, I believe market is rewarding Nebius by using Databricks as a valuation comparison but this is misconstrued.
The similarities the market sees is that both companies have a full inference stack with Databrick provides LLM inference through API by the token through Databricks Mosaic (32, 33) and with Databricks currently at 6.9B ARR (34) and Nebius targetting a 7B-9B ARR by EOY. The big mistake here is that NBIS revenue is for SaaS plus infrastructure while Databricks is SaaS without infrastructure.
Databricks does not need to own GPUs, power contracts, or data centers to recognize their revenue. Databrick's 74% gross margin (34) is even better than Anthropic's 80% gross margins because Databricks doesn't have to pay for model training like Nebius because it uses Open Source Model. At the same time it's compute cost is like Anthropic's: accounted for in the gross margins and not subtract off afterwards as depreciation.
Of course Databricks does more than Managed Inference for Open Source and has Lakehouse as well. Databricks is also one step ahead with Managed Agentic Agents (36).
Competitors
There are way more Neoclouds than finX is aware of because many successful companies stay private nowadays longer and the field of AI Platforms is still young. Notable competitors are hyperscalers, Cerebras, FireworksAI, SambaNova, FireworksAI, TogetherAI, Baseten, Modal, Databricks, Coreweave and weaker competitors like Cloudflare, Novita, DeepInfra, Parasail, Scaleway, Lambda, Mistral, Cohere, AI21 Labs, OctoAI.
Every Neocloud likes to cherry-pick benchmarks for best inference speeds that make them look the best but we should select open source (OS) model first and then look at results. The models we select are:
1. GLM-5.2 Max - the most powerful OS model today.
2. DeepSeek V4 Pro - the most deployed in production usage near-frontier OS model model today.
3. gpt-oss-120b - the OS model with most benchmark submissions
The results are pictured in the same order as I listed them above and Nebius is below average.
artificialanalysis.ai/models/glm-5-2…
artificialanalysis.ai/models/deepsee…
artificialanalysis.ai/models/gpt-oss…
Granted Nebius acquired EigenAI and Clarifai which look good on certain 1 dimensional benchmarks. These startups are optimizing latency benchmarks for buyouts and not real economics. For startups not serving large number of customers, they can juice their benchmark numbers by reducing batch size, over provisioning GPUs, keep KV cache longer in HBM, increase decoding budget, tune for latency over GPU utilization.
Moving Fast
Fireworks came out the other day with Managed Model Tuning of Open Source Models called Fireworks RFT (38). Basically for those familiar with Nebius lingo, it's Token Factory but Model Factory.
I expect Nebius to follow suit in 6-12 months. But by that time, other AI Platforms Databricks, Coreweave will have already released through developer documentation instead of marketing fanfare.
Obviously then there are obvious risks of OpenAI and Anthropic dominating and SpaceX DCs which might be cost competitive with $BE fuel cell sites.
In all fairness, if you are reading this for the first time and want the bull case, these accounts have done a great job laying out the $NBIS bull case:
1. @daniel_koss
2. @itsalasdairmann
3. @MB_Hogan
4. @babyfolio
5. @EndicottInvests
I don't own any $NBIS but some $IREN whales do own a fair amount of $NBIS and have done well with $IREN in 2025 and $NBIS in 2026.
There's no doubt that in H1 2026, $NBIS has had the best stock performance among publicly traded Neoclouds, with concrete deliverables in 399m Q1 revenue on 82% QoQ growth (1) and 40% downpayment from MSFT (2). At 11-15m/MW-yr, 399m quarterly revenue equates to 106MW-145MW of build out depending on blend GPUs and IaaS-PaaS-SaaS. This buildout is still early and not bottlenecked by power. However, hot stock performance for one time frame doesn't determine the whole trajectory; an example is $SMCI vs $DELL and $HPE. What will de-risk $NBIS for me is completion of their Vineland, NJ datacenter in 2026 or early 2027.
I will go over two main topics: Power and Software.
1. Power
Nebius has more than 4GWs of secured power with a strategy of diversifying between colocation and self-owned over many sites to manage risk. Nebius refers to secure and contracted power as either PPA from the energy producer or interconnection agreement (IA) which is final approval needed for grid buildout. Nebius' BTM sites like Vineland, NJ do not need IA and also bypass the grid buildout time.
Not all Nebius sites are equal so I'll go through starting from the ones that are the best in terms of time to power without delays.
Mäntsälä
Nebius's Mäntsälä site is 75MW of active power (11). This datacenter from Nebius's Yandex days is grid connected, greenfield, and self-owned (11). The Finland datacenter boast a PUE of 1.13 (11) comparable with IREN's Prince George datacenter at 1.1 PUE (12), both taking advantage of the climate.
Iceland
10MW of colocation lock and loaded (26).
Kansas City, Missouri
5GW of Colocation from Patmos lock and loaded, potential expansion to 40MW (29).
Israel
Israel my favorite $NBIS site in terms of guaranteed path to power with no delays. Nebius signed a datacenter lease for 80MW in Modi'in in the Masmiyya and Beit Shemesh districts with expansion to 286MW (23). Some of the clusters will be used for Israel's national super computer (24) and Beit Shemesh is an important military district, previously target of Iran's military strikes but now de-risked as the Iran Wwar has ended (25).
Independence
Of Nebius's large sites in North America, I like the potential of Independence the best. It cleverly uses reopening an old power plant built in 1958 and closed in 2020 near by to provide power. The natural gas-fueled Blue Valley Power Plant is very close by Nebius' designated site so grid buildout requires less high voltage components.
This is similar to Elon's colossus site which uses a mix of reopened power plant and importing gas turbines from deconstruction power plants from abroad (22).
UK
Ark Data Centres colocation, 16MW agreed with expansion plans to 52MW (30).
Vineland
Nebius current active buildout is a Colcoation site in Vineland, NJ with "Nebius, becoming a DataOne tenant for the next 10 years" (7). From Nebius' 20-F at the end of 2024 (9) and F-3 mid 2025 (10), "The New Jersey site is a phased development scalable up to 300 MW, with initial capacity expected to be available in the second half of 2025." However, we are entering into the second half of 2026 and Vineland has just switched out their power supply from Bergen cruiseship engines with Skylea carbon capture (13) to $BE fuel cells (14).
Nebius is paying 2.6B for 328MW of $BE Fuel Cells as a Service for 10 years with 250MW of guaranteed capacity and 78MW of installed system capacity (14). Like all BTM power sources, redundancy factor is needed which is the 78MW and only 250MW goes to the GPU and Aux Loads. This is 10.4m/MW for 10 years or 1.04m/MW-yr and doesn't even include the natural gas (17). To put in perspective, the entire datacenter cost buildout 15m/MW for 20 years with 3-4m retorfit cost if a site has grid connected power.
There is a concern whether or not Nebius will get the $BE fuel cells this year as they are sold out for 2026 (16). This is not a concern and the fuel cells themselves will be delivered this year because Oracle encountered a delay for their NM site and $BE is rerouting the fuel cells from Oracle to Nebius (15).
As stated in the 6-K, Nebius may provide "alternative capacity" (31) for it's Microsoft contract while Vineland is under construction. This will work for several tranches but will not work forever as Vineland is 3.48B ARR and almost half of $NBIS 7-9B 2026 ARR guidance.
Lappeenranta
Nebius has labled Lappeenranta, Finland as a 310MW greenfield site. Lappeenaranta is a greenfield site belonging to Polarnode and Nebius is renting it as Colocation (18). @TheUglyBuckling has full details on this site: x.com/TheUglyBucklin…
Bethune, France
@InvestNorthwise does a good job of covering this one here: northwiseproject.com/nbis-bethune-f….
Birmingham, Alabama
300MW of brownfield development that got permitted right before the city passed it's datacenter moratorium. This site has broken ground and has a decent chance to completion as a BTM site despite lawsuit from two homeowners who are within 1000ft of the datacenter (27).
The major blocker from Birmingham succeeding as a grid connected site is that the permit to build a substation and switching station was denied and final (28). The workaround to this is obviously just like Vineland $BE fuel cells. This is why I see Vineland, NJ as a pivotal site for Nebius as alot of their sites' contingency plan is $BE fuel cells so Vineland is the litmus test for their ability to execute on $BE fuel cell BTM sites.
Pennsylvania
Nebius most recent site announcement is their 1.2GW Pennsylvania site which they revealed to Analyst is Pennsylvania is grid connected (3). By end of 2027, Pennsylvania will have 250MW-350MW with each year adding 300MW (5). From cross referencing total capacity, power ramp, and acreage of site, Nebius's Pennsylvania site is most likely Lower Mount Bethel Technology Center” in Northampton County (6). If so, Nebius has secured the power purchase agreement and still needs to get the interconnection agreement which can be delayed based on time to build transmission, substations, and grid. Regulations and local anti-datacenter activism is dealt with in parallel to optimize time to power.
Good news is PJM is reforming it's interconnection queue to be first-come first-serve (19). Bad news is despite that, PJM is one of the most congested buildouts from a supply chain perspective and "PJM Interconnection data – but the biggest delays are no longer occurring within the interconnection queue" (20).
"The bottleneck has shifted downstream. Transmission buildouts, substation capacity, and strained supply chains are now the primary obstacles to energizing projects. According to PJM, projects spent an average of more than three years reaching an interconnection service agreement, and another four years waiting to come online after approval" (20).
2. Software
A large part of Nebius market cap is attributed software. Although Nebius is way ahead of IREN in software, I believe market is rewarding Nebius by using Databricks as a valuation comparison but this is misconstrued.
The similarities the market sees is that both companies have a full inference stack with Databrick provides LLM inference through API by the token through Databricks Mosaic (32, 33) and with Databricks currently at 6.9B ARR (34) and Nebius targetting a 7B-9B ARR by EOY. The big mistake here is that NBIS revenue is for SaaS plus infrastructure while Databricks is SaaS without infrastructure.
Databricks does not need to own GPUs, power contracts, or data centers to recognize their revenue. Databrick's 74% gross margin (34) is even better than Anthropic's 80% gross margins because Databricks doesn't have to pay for model training like Nebius because it uses Open Source Model. At the same time it's compute cost is like Anthropic's: accounted for in the gross margins and not subtract off afterwards as depreciation.
Of course Databricks does more than Managed Inference for Open Source and has Lakehouse as well. Databricks is also one step ahead with Managed Agentic Agents (36).
Competitors
There are way more Neoclouds than finX is aware of because many successful companies stay private nowadays longer and the field of AI Platforms is still young. Notable competitors are hyperscalers, Cerebras, FireworksAI, SambaNova, FireworksAI, TogetherAI, Baseten, Modal, Databricks, Coreweave and weaker competitors like Cloudflare, Novita, DeepInfra, Parasail, Scaleway, Lambda, Mistral, Cohere, AI21 Labs, OctoAI.
Every Neocloud likes to cherry-pick benchmarks for best inference speeds that make them look the best but we should select open source (OS) model first and then look at results. The models we select are:
1. GLM-5.2 Max - the most powerful OS model today.
2. DeepSeek V4 Pro - the most deployed in production usage near-frontier OS model model today.
3. gpt-oss-120b - the OS model with most benchmark submissions
The results are pictured in the same order as I listed them above and Nebius is below average.
artificialanalysis.ai/models/glm-5-2…
artificialanalysis.ai/models/deepsee…
artificialanalysis.ai/models/gpt-oss…
Granted Nebius acquired EigenAI and Clarifai which look good on certain 1 dimensional benchmarks. These startups are optimizing latency benchmarks for buyouts and not real economics. For startups not serving large number of customers, they can juice their benchmark numbers by reducing batch size, over provisioning GPUs, keep KV cache longer in HBM, increase decoding budget, tune for latency over GPU utilization.
Moving Fast
Fireworks came out the other day with Managed Model Tuning of Open Source Models called Fireworks RFT (38). Basically for those familiar with Nebius lingo, it's Token Factory but Model Factory.
I expect Nebius to follow suit in 6-12 months. But by that time, other AI Platforms Databricks, Coreweave will have already released through developer documentation instead of marketing fanfare.
Obviously then there are obvious risks of OpenAI and Anthropic dominating and SpaceX DCs which might be cost competitive with $BE fuel cell sites.
17
@GavinSBaker
Jim Liu@jiahanjimliu
$IREN / $NBIS: Asset Lite Model, Vertical Integration, DSX OS
Horizontal vs Vertical Integration
Last year, NBIS bulls were pushing the vertical integrating narrative to which I said that owning the GPUs doesn't make vertical integration (3) because the vertical integration is handled by Nvidia libraries that sit between the GPU and higher level software, most noticeably CUDA, NCCL.
Now NBIS bulls have change the narrative to horizontal integration where they don't need to own the GPUs and can scale by licensing up their system architecture and software stack (2). All Neoclouds base their system architecture around DSX reference design (1), so it's really the software stack we are talking about here.
The new NBIS narrative of horizontal integration is correct while vertical integration is a narrative that NBIS made up to seem superior. Today's NBIS announcement made it clear that vertical integration has nothing to do with owning the GPUs design by Nvidia and everything to do with CUDA, NCCL.
Asset Lite
Yes, this is not the exact same as colocation. Asset lite is colocation plus switching whose balance sheet the GPU and networking gear are on.
This is a fine business model for NBIS. However, its nothing revolutionary as FireworksAI, TogetherAI, Baseten, Modal have been doing asset lite for years. In fact Fireworks, TogetherAI are AI Platforms for IREN bare metal.
NBIS bulls are masters of narrative and say that FireworksAI and TogetherAI don't have vertical integration when they do this with IREN. Now NBIS is doing this themselves but labeling it as Asset Lite instead of horizontal integration.
Ironically, NBIS and IREN were both on the asset side of this model as they provided raw compute for MSFT Azure AI Platform. NBIS even signed another contract as a compute provider for Meta too. NBIS likes to spin the narrative that they are providing software for all their customers but it's very common to add on Kubernetes on top of bare metal as bare metal+. This is very close to raw compute at the end of the day.
DSX OS
Some NBIS bulls are now trying to narrate that DSX is Nvidia specs on how to build a datacenter and doesn't come with software (4). This could not be further than the truth.
Straight from Nvidia: "As part of the NVIDIA DSX platform, DSX OS delivers composable components for lifecycle management, runtime consistency, health automation, resiliency, multi-tenant operations, and AI platform services" (5).
In fact, Nvidia acquired RunAI, an Israeli based startup, for 700m which specializes in GPU Orchestration (6). Today you can install RunAI for free (7) as long as you have Kubernetes setup. For IREN, Mirantis can easily integrate RunAI. Some NBIS bulls are saying NBIS has the best stack (9) but really RunAI is the standard which most Neoclouds and Hyperscalers base their own orchestration around. RunAI has stats to back up their performance (10) and now that Nvidia engineers have had 2 addition years to work on it, RunAI is the industry standard base code.
IREN Pioneering DSX OS
IREN will be the first to pioneer an fully open source integration of DSX OS via Mirantis (11). This is why Nvidia setup IREN up to acquire Mirantis to begin with.
Previously we have hyperscalers and Neoclouds integrating DSX OS and making it proprietary in order to stick their customers. Now Nvidia is tasking IREN-Mirantis to have the full integration open sourced (11). Yes you head it right! The AI Platform open sourced. Granted the DSX OS doesn't include inference optimizations but you can bet Nvidia is going to have open source inference optimizations for Nemotron. As Chinese open source companies like Deepseek expand from Model company to Cloud (12), you can get they will release open source inference optimizations along with open source models.
Conclusion
The ground is shifting, keep watch of the Neocloud industry as it developer. NBIS is great at framing and narration, but Nvidia leads the industry including AI Cloud Platform and IREN is here to participate.
Horizontal vs Vertical Integration
Last year, NBIS bulls were pushing the vertical integrating narrative to which I said that owning the GPUs doesn't make vertical integration (3) because the vertical integration is handled by Nvidia libraries that sit between the GPU and higher level software, most noticeably CUDA, NCCL.
Now NBIS bulls have change the narrative to horizontal integration where they don't need to own the GPUs and can scale by licensing up their system architecture and software stack (2). All Neoclouds base their system architecture around DSX reference design (1), so it's really the software stack we are talking about here.
The new NBIS narrative of horizontal integration is correct while vertical integration is a narrative that NBIS made up to seem superior. Today's NBIS announcement made it clear that vertical integration has nothing to do with owning the GPUs design by Nvidia and everything to do with CUDA, NCCL.
Asset Lite
Yes, this is not the exact same as colocation. Asset lite is colocation plus switching whose balance sheet the GPU and networking gear are on.
This is a fine business model for NBIS. However, its nothing revolutionary as FireworksAI, TogetherAI, Baseten, Modal have been doing asset lite for years. In fact Fireworks, TogetherAI are AI Platforms for IREN bare metal.
NBIS bulls are masters of narrative and say that FireworksAI and TogetherAI don't have vertical integration when they do this with IREN. Now NBIS is doing this themselves but labeling it as Asset Lite instead of horizontal integration.
Ironically, NBIS and IREN were both on the asset side of this model as they provided raw compute for MSFT Azure AI Platform. NBIS even signed another contract as a compute provider for Meta too. NBIS likes to spin the narrative that they are providing software for all their customers but it's very common to add on Kubernetes on top of bare metal as bare metal+. This is very close to raw compute at the end of the day.
DSX OS
Some NBIS bulls are now trying to narrate that DSX is Nvidia specs on how to build a datacenter and doesn't come with software (4). This could not be further than the truth.
Straight from Nvidia: "As part of the NVIDIA DSX platform, DSX OS delivers composable components for lifecycle management, runtime consistency, health automation, resiliency, multi-tenant operations, and AI platform services" (5).
In fact, Nvidia acquired RunAI, an Israeli based startup, for 700m which specializes in GPU Orchestration (6). Today you can install RunAI for free (7) as long as you have Kubernetes setup. For IREN, Mirantis can easily integrate RunAI. Some NBIS bulls are saying NBIS has the best stack (9) but really RunAI is the standard which most Neoclouds and Hyperscalers base their own orchestration around. RunAI has stats to back up their performance (10) and now that Nvidia engineers have had 2 addition years to work on it, RunAI is the industry standard base code.
IREN Pioneering DSX OS
IREN will be the first to pioneer an fully open source integration of DSX OS via Mirantis (11). This is why Nvidia setup IREN up to acquire Mirantis to begin with.
Previously we have hyperscalers and Neoclouds integrating DSX OS and making it proprietary in order to stick their customers. Now Nvidia is tasking IREN-Mirantis to have the full integration open sourced (11). Yes you head it right! The AI Platform open sourced. Granted the DSX OS doesn't include inference optimizations but you can bet Nvidia is going to have open source inference optimizations for Nemotron. As Chinese open source companies like Deepseek expand from Model company to Cloud (12), you can get they will release open source inference optimizations along with open source models.
Conclusion
The ground is shifting, keep watch of the Neocloud industry as it developer. NBIS is great at framing and narration, but Nvidia leads the industry including AI Cloud Platform and IREN is here to participate.
18
@GavinSBaker
Jim Liu@jiahanjimliu
Natural Gas Shortage Starting H2 2028
Matthew Smith, CIO of Chronometer Partners, has done a bottoms up modeling of every asset in the US gas system and the results counter intuitive.
The conclusion is that datacenters relying on BTM Natural Gas could see natural gas cost become 20-30% of their cost of compute. This will hit AI compute much like memory shortage which nothing happened then all at once.
The Counter Intuition
Right now, natural gas is so abundant because of the shale revolution and the US has massive excess of natural gas because it comes with oil fracking. However, the shale revolution is a one time breakthrough that unlocks much of the natural gas production capacity in the US. Additional natural gas investments will not be a natural byproduct of oil fracking and the bottleneck will come from upstream natural gas infrastructure.
A significant portion of excess US natural gas production has been committed in long term contracts to US allies as part of US foreign policy. 1/3 of all the LNG in the world comes from the US. To pull the plug back for foregin exports would not just be breaking long term contracts but also be a matter of foreign policy.
Much of the projected AI buildout will be BTM Natural Gas. SMRs are almost certain to face delays if not outright commercialization issues. SMRs are more complex than Space datacenters. Large scale nuclear takes many years of advance planning and construction timelines even in China.
The Constraints
The real constraint is upstream gas deliverability. This includes hydraulic fracturing equipment, compressor stations or transmission pipelines.
Matthew predicts natural gas prices to triple by 2029.
Winners
1. Datacenter buildouts with renewable grid energy - $IREN, $GOOG, $AMZN
2. Space DCs - $SPCX
3. Large Nuclear Reactors - $CCJ, $BWXT
4. Natural Gas Producers with Real Inventory - $EXE, $CRK, $RRC
5. BTM Sites that Secure Natural Gas First - $NUAI
Loosers
1. Natural Gas BTM buildouts that secure natural gas as an auxilary step.
2. Any Datacenter Site in PJM - PJM grid is highly reliant on natural gas. Particularly, Pennsylvania, New Jersey, and Maryland.
3. Gas Turbine Makers / Fuel Cells - $CAT, $AGX, $BE, $FCEL. There will not be enough gas to fuel the expanded production capacity these companies plan for.
4. Consumers - consumers need protection against AI datacenters bidding up natural gas prices.
Conclusion
Every commodity is a commodity until it hits inflection point. While HBM is not a commodity, DRAM has been a commodity for over 20 years until last year. Once production hits physical limits, commodity prices go stratospheric.
Matthew Smith, CIO of Chronometer Partners, has done a bottoms up modeling of every asset in the US gas system and the results counter intuitive.
The conclusion is that datacenters relying on BTM Natural Gas could see natural gas cost become 20-30% of their cost of compute. This will hit AI compute much like memory shortage which nothing happened then all at once.
The Counter Intuition
Right now, natural gas is so abundant because of the shale revolution and the US has massive excess of natural gas because it comes with oil fracking. However, the shale revolution is a one time breakthrough that unlocks much of the natural gas production capacity in the US. Additional natural gas investments will not be a natural byproduct of oil fracking and the bottleneck will come from upstream natural gas infrastructure.
A significant portion of excess US natural gas production has been committed in long term contracts to US allies as part of US foreign policy. 1/3 of all the LNG in the world comes from the US. To pull the plug back for foregin exports would not just be breaking long term contracts but also be a matter of foreign policy.
Much of the projected AI buildout will be BTM Natural Gas. SMRs are almost certain to face delays if not outright commercialization issues. SMRs are more complex than Space datacenters. Large scale nuclear takes many years of advance planning and construction timelines even in China.
The Constraints
The real constraint is upstream gas deliverability. This includes hydraulic fracturing equipment, compressor stations or transmission pipelines.
Matthew predicts natural gas prices to triple by 2029.
Winners
1. Datacenter buildouts with renewable grid energy - $IREN, $GOOG, $AMZN
2. Space DCs - $SPCX
3. Large Nuclear Reactors - $CCJ, $BWXT
4. Natural Gas Producers with Real Inventory - $EXE, $CRK, $RRC
5. BTM Sites that Secure Natural Gas First - $NUAI
Loosers
1. Natural Gas BTM buildouts that secure natural gas as an auxilary step.
2. Any Datacenter Site in PJM - PJM grid is highly reliant on natural gas. Particularly, Pennsylvania, New Jersey, and Maryland.
3. Gas Turbine Makers / Fuel Cells - $CAT, $AGX, $BE, $FCEL. There will not be enough gas to fuel the expanded production capacity these companies plan for.
4. Consumers - consumers need protection against AI datacenters bidding up natural gas prices.
Conclusion
Every commodity is a commodity until it hits inflection point. While HBM is not a commodity, DRAM has been a commodity for over 20 years until last year. Once production hits physical limits, commodity prices go stratospheric.