INSEAD 19D | Ex-management consultant | Exploiting the disconnect between secular trends and asset pricing
Dubai
Joined September 2009
- Tweets1.4K
- Following705
- Followers1.4K
- Likes18.3K
$IREN ¡Hala Madrid!
IREN is advancing sovereign AI in Europe, bringing together its European data center capacity and managed cloud services for customers seeking more choice over where their AI workloads run and how their data and infrastructure are managed.
Announced today at NVIDIA GTC Berlin, IREN has established IREN Europe and appointed Ángel Corcóstegui as Chair of the newly established board of directors.
IREN also secured the civil works license for its planned 300MW campus in Badajoz, Spain. Following the completion of its acquisition of Mirantis, IREN’s software capabilities include workload isolation and operational controls that can help customers address data residency, security and governance requirements.
Read more: iren.gcs-web.com/static-file…
Highly recommended read for $IREN shareholders.
1) Trying again.
TLDR: Everyone can win. Open-source, frontier labs, infrastructure, applications and most importantly the American people.
I understand all of the jitters about Anthropic losing share, real rates, spreads and second derivatives. Still think the most likely future outcome is a hybrid, multi-model future where almost everyone wins, open-weight tokens are likely the overwhelming majority of tokens, the dominant frontier model companies still do well albeit perhaps not the outcome that seemed possible in May and AI infrastructure/applications benefit immensely from the increased fragmentation, competition and lower margins at the model layer.
Lower margins at the model layer mechanically drive higher compute demand as long as someone is willing to pay for the training. And Jensen is obviously willing to be the “trainer of last resort.” As below-market compute contracts signed in 24/25 roll-off and reprice higher, cloud and neocloud revenue should continue to accelerate which will help address financing fears; I think we may be near peak external financing for this buildout. Noteworthy that using data from 3Fourteen Research, B200 clusters will generate a 37% IRR *if* their rental price declines 40% per year. Prices are still going up, but that IRR math is too ridiculous to print. 37% IRRs available at scale should get funded.
Re: this weekends controversy about MOUs, Anthropic’s $413 billion of real contractual commitments sounds scary. Reality is that it works out to roughly $31 billion of annual cloud spend, which is easily internally financed given their reported ARR, and $161 billion of spend at Broadcom which will be financed by Broadcom and private credit.
Continue to believe fragmentation at the model layer is the best outcome for ordinary Americans and humans everywhere. As I have said before, do not trust any 5-20 individuals to make the best possible decisions about the direction of this technology. I want as many AIs as possible to maximize the odds that they encompass the diversity of human values across the world and richness of human variation. And I want an AI that shares my own personal values. Safety, freedom and the maximization of human agency through democratization, distribution and diffusion vs. centralization. History has spoken on the dangers of the latter path.
Finally, from my perspective anything that slows this buildout down is net positive over the long-term as the biggest risk and worst outcome is an *actual* overbuild. We are structurally short watts, wafers and increasingly capital. That is all good. God forbid we were still in ZIRP, TSM ramped capex to 3x their current plan and we weren’t short turbine blades: there would almost certainly be an overbuild and a bubble. That would be terrible. Slower and smoother for longer is better both for investors and for the world. A terrible cyber event is the most plausible near term bear case IMHO, but AI cyberdefense ramping quickly and I was happy to hear Sholto say that cyber probably becomes defense dominant in the near future.
Longer, threaded version follows:
Here’s some math for you based on the economics $IREN is discussing today.
Per 1 MW of IT load:
Revenue: ~$25M/year
Total operating expenses: ~($5M)
Adjusted EBITDA: ~$20M
~80% margin
D&A: ~($8M)
~$7M GPUs @ $35M capex depreciated over 5 years
~$1M data center @ $15M capex depreciated over 15 years
EBIT: ~$12M
Interest expense: ~($3M)
90% of GPU capex @ ~9% rate
Pre tax earnings: ~$9M per MW per year
Close your shorts before it’s too late.
Neocloud Math: JP Morgan seems to believe that $5B of D&A in FY 2030 on $80 to $100B of capital employed (2.1 GW) is proper. $IREN
$IREN What you are seeing and hearing from frontier model CEOs is partly coordinated theatre to put themselves in a better strategic position in the AI race.
Their safety concerns can be genuine. But slowing down frontier development also makes a lot of economic sense for them.
Here is why:
Current models are already capable enough to drive massive adoption for at least the next few years.
Even if model progress stopped today, demand for tokens would still keep growing very fast.
Think about what current models can already do. You don’t need another 10 generations of models before businesses start adopting this stuff.
Any rational profit seeking company that can materially increase employee productivity using AI will eventually want to do it.
So the problem frontier labs have today is not that their models are not good enough.
The problem is that they are spending insane amounts of money racing each other to make them even better, while the infrastructure available to serve demand for existing capabilities is already constrained, and will probably remain constrained for at least the next 2 to 3 years.
If current capabilities are already enough to drive the next 2 to 3 years of adoption, pushing the frontier faster does not create more near term demand.
What it definitely does increase is training capex.
If all frontier labs slow the pace of model development, they can lengthen their training cycles and move more of their compute toward inference.
Training is where they burn huge amounts of capital. Inference is where they actually make money.
So instead of constantly spending billions trying to leapfrog each other, they can spend more time monetizing the intelligence they already built.
The obvious problem is the prisoner’s dilemma.
No single lab can slow down if everyone else keeps pushing.
Every player has an incentive to cheat, because whoever continues pushing while everyone else slows down could gain a huge advantage.
So if you really want to slow frontier development, you need everyone operating under the same rules.
And because this race is not only happening in the US, eventually China has to be part of that framework too.
Frontier companies are also reaching the point where they need to justify their economics to investors.
It is difficult to justify huge valuations if the answer to every additional dollar of revenue is another massive increase in training capex.
One way to improve those economics is simple.
Slow the training race.
Lengthen model cycles.
Move more compute toward inference.
Grow revenues faster than training capex.
Improve margins.
You basically stop giving away so much of the economic surplus from better models to the consumer every few months and spend more time monetizing what you already have.
And paradoxically, I think this could actually be bullish for AI long term.
A market where every player is forced to spend more and more just to avoid falling behind is not necessarily a healthy market.
At some point, investors will ask for returns.
If these companies can slow the training arms race, improve margins and prove that AI can generate massive cash flows before the next capex cycle, you create a much more sustainable environment for the industry.
That means more capital will be willing to stay in AI for longer.
So slower frontier progress in the short term could actually make the AI investment cycle stronger and longer.
And what does this mean for neoclouds?
If inference demand keeps growing faster than infrastructure supply, the capacity simply moves from training to inference.
The important question for neoclouds is not whether frontier labs train new models every 6 months or every 12 months.
The important question is whether total demand for compute keeps growing.
Given how early we still are in AI adoption, I think it does.
$IREN This is why IREN isn’t rushing to sign deals 12–18 months in advance, with Microsoft being the exception.
The closer an uncontracted MW gets to its delivery date, the larger the pool of credible bidders becomes, and the more valuable that MW gets.
At some point, bidders will understand this dynamic too. Once they realize that waiting means competing against a much larger buyer pool at potentially much higher prices, they’ll start bidding for MWs 12–18 months out just to secure capacity and make sure they are able to execute their AI roadmaps.
The price of each MW will explode.
2. The forward-contracting question has flipped 180 degrees. Not long ago it was "wen deal". This week: how long can you hold your capacity available. The structural point is being understood: only half a dozen companies in the world can contract capacity 12-18 months ahead of commissioning. Every month closer to delivery, that universe expands, and capacity allocation becomes a strategic and economic decision rather than a funding necessity.
I increased my $IREN position by 10% over the past month, then added another 20% on Friday after the earnings call.
Forget all the BS. You really only need to understand 3 data points to see where this is headed:
1- Revenue per MW is going parabolic, and IREN still has 5GW+ to monetize.
2- Despite a few hiccups, the team is executing beautifully: diversifying the customer mix, financing efficiently, deepening the NVIDIA relationship and credentials, acquiring Mirantis, delivering Horizon 1 to Microsoft, etc.
3- Most importantly, IREN can theoretically finance future growth with very little equity, and potentially zero equity once DC-level financing is layered in. They basically spelled it out on the call:
A. Customer prepayments are now covering ~50% of GPU capex.
B. IREN is financing 90% of GPU capex with debt. With a 50% customer prepayment, debt + prepayment equals 140% of GPU capex, leaving an amount equal to 40% of GPU cost available to help fund the data center.
C. GPUs are roughly two-thirds of the fully loaded GPU + DC cost, meaning DC capex is roughly 50% of GPU capex.
So for every $150 of total capex, IREN could theoretically need to contribute only $10 of equity, or ~7%, and that’s before tapping any DC or corporate-level financing.
At a ~$14B MC, with >$4B of ARR expected by the end of December and a plausible path to >$15B ARR by the end of 2027, I struggle to find a better asymmetric bet.
Doubled my $IREN position this week. I think a re-rating is imminent. My gut says they’ll be ordering a second, even bigger batch of Blackwells soon. That’s when the real fireworks begin IMO.
Jacques B. retweeted
$IREN the clipped highlighted part caught my attention:
Asked Gemini Pro to breakdown the DSX infrastructure:
Here is a breakdown of what the infrastructure entails and why it represents a shift in the industry:
1. Full-Stack Vertical IntegrationDSX is designed to unify the entire physical and digital environment of an AI data center. The architecture integrates:
Accelerated Compute: The high-density GPU clusters required for training and inference.
High-Speed Networking: The complex internal fabric needed so thousands of GPUs can communicate without bottlenecks.
Physical Infrastructure: Standardized layouts for advanced power delivery systems and specialized liquid cooling solutions.
Software: The operating layer that dynamically orchestrates AI workloads to maximize energy efficiency and intelligence-per-watt.
2. Solving the Power BottleneckThe primary constraint on modern AI expansion is no longer just manufacturing chips; it is the global shortage of utility-scale power, grid interconnection queues, and specialized data center facilities. The DSX architecture is engineered to be deployed in partnership with specialized infrastructure operators who already control large land portfolios and gigawatt-level power contracts.
3. Repeatable "Turnkey" DeploymentBy establishing a de facto standard for how an AI factory should be built from the ground up, DSX removes the guesswork for vendors, integrators, and colocation providers. It allows facilities to be planned and executed as a single, cohesive unit, dramatically accelerating the time it takes to bring gigawatt-scale AI compute online compared to custom-building each facility.
In short, DSX represents a structural shift in the AI hardware market: moving away from procuring GPUs in isolation, and toward deploying fully standardized, turnkey AI factories at a massive scale.
This quoted post is unavailable.
Jacques B. retweeted
There is no larger long-term strategic move than this — NVIDIA joins forces with $IREN to build the flagship AI factory deployment for the DSX architecture
The market will continue to repeatedly reinterpret the deeper intent and long-term objectives behind the partnership between NVIDIA and IREN.
On May 7, 2026, IREN’s CEO reposted NVIDIA’s official announcement on X regarding the partnership between the two companies: NVIDIA and IREN Limited today announced a strategic partnership to accelerate the deployment of next-generation AI infrastructure.
NVIDIA announcement
nvidianews.nvidia.com/news/n…
At the same time, IREN also released another announcement on its own website: IREN signs a US$3.4 billion AI cloud services agreement with NVIDIA.
IREN announcement
iren.gcs-web.com/static-file…
The two announcements, each emphasizing different aspects of the cooperation, carry extremely significant implications.
First, after careful verification, this is the first time NVIDIA has sought external compute leasing. There are three major turning points in industry development embedded in this move.
A reversal of roles: NVIDIA becomes a “major external compute customer” for the first time
In the past, NVIDIA’s relationship with infrastructure companies was almost always centered around “selling hardware” or “borrowing hyperscaler data centers for DGX Cloud.” But in this US$3.4 billion agreement with IREN, NVIDIA is, for the first time in its history, leasing third-party compute capacity at large scale and on a long-term basis as a customer, for use by its own AI research teams. This kind of “reverse leasing” is unprecedented for NVIDIA in both scale and nature.
The selective external exposure of its most core secrets: this point carries the deepest implications
For a long time, NVIDIA has insisted on keeping its most critical R&D work — chip design, driver optimization, and large-model training — inside its self-built supercomputers such as Selene and Eos, creating a closed loop of “building the shovels and mining with them itself.” But this time, outsourcing a 60MW research workload to an external data center is highly significant. It signals that compute-chip R&D is beginning to transition toward external collaboration.
The first opening of stack management: introducing Mirantis to manage NVIDIA’s internal R&D clusters
Previously, NVIDIA’s internal cluster management was handled entirely by its own engineering teams. But under this agreement, NVIDIA is for the first time allowing third-party management, bringing in Mirantis to participate in cluster orchestration and operations. This also signals a transformation in NVIDIA’s latest compute architecture R&D approach — beginning to “strengthen external collaboration” for lower-level operational work such as server cooling, restarts, and Kubernetes configuration.
As the ability of individual GPU chips to increase computing performance gradually approaches physical and engineering limits, the next phase of AI compute advancement is shifting from “single-chip performance competition” to “system-level scalability competition.” This is NVIDIA’s direction of transformation.
The primary paths for the next stage of AI compute improvement include: GPU clustering, high-speed interconnects, rack-scale computing, and data-center-level coordination. This requires GPU manufacturers (NVIDIA), data center designers/builders/operators (IREN), and supercluster operating systems (Mirantis) to jointly collaborate on development.
What they are developing is precisely the NVIDIA DSX architecture referenced in the NVIDIA-IREN partnership announcement. And IREN’s hyperscale SW site in Texas is becoming the flagship deployment location for NVIDIA’s DSX architecture. This is absolutely not a simple narrative of NVIDIA investing in a company and becoming a shareholder.
For the world’s leading company that holds the core secrets of AI compute chip R&D, this is not a trivial matter.
From NVIDIA’s perspective, there appear to be many potential partners, such as CoreWeave, Nebius, Oracle, Microsoft Azure, Amazon Web Services, and Crusoe, and NVIDIA has already invested in or partnered with these firms before. But why did it choose IREN for this most important transformation?
Because IREN possesses too many things that are uniquely its own:
Multiple GW-scale single sites with secured long-term power supply
Grid interaction capabilities
Vertical integration
Ultra-long-term site planning and abundant land supply
Green energy
Acting as its own design-and-build general contractor
Long-term accumulation of data center operational experience
Advanced design and technical capabilities
Compared with the companies above that NVIDIA has already partnered with, even if IREN temporarily lacked software capabilities, NVIDIA was still willing to wait until IREN acquired a software company before announcing this deep cooperation. Moreover, Mirantis has long been one of the three software companies that have collaborated with NVIDIA for many years. It is highly possible that NVIDIA itself played the role of connector behind IREN’s acquisition.
NVIDIA is transforming toward system-level compute scaling and building an AI factory template. In the future, the products it sells may no longer simply be GPU chips, but complete racks, clusters, or even entire AI factories.
That inevitably requires standardized data centers in order to guarantee performance, compatibility, scalability, and token efficiency.
What NVIDIA needs are facilities with massive long-term secured power supply, land, GW-scale campuses, HPC DNA, rapid construction capability, neutrality, automated scheduling capability, workload routing, GPU virtualization, fault recovery, and cluster operating systems capable of distributed training management.
At present, IREN is the only company that possesses all of these elements simultaneously.
What they are trying to build is the industrial standard for the next phase of the AI industry.
The greatest companies do not merely participate in industries — they define the standards.
From this perspective, there is no larger strategic theme than this one.
Selling compute capacity to hyperscalers, partnering with Anthropic, or developing new sovereign AI businesses are all important, but none compare with this.
The deeper meaning of last week’s announcement will require time for the market to fully interpret and understand. I believe I have already analyzed this trend relatively clearly.
This move by NVIDIA and IREN, once executed successfully, could once again widen the gap between the NVIDIA ecosystem and Google just as Google had begun catching up — and it carries major implications for the entire AI industry.
$IREN I don’t think the Mirantis acquisition and the NVIDIA partnership should be viewed as separate events.
I think IREN is positioning itself as one of the physical and operational deployment layers for AI factories.
IREN brings the scarce physical layer: land, power, data centers, construction speed, and operational execution.
NVIDIA brings the compute layer: GPUs, networking, systems architecture, AI factory design, customer pull, and ecosystem credibility.
Mirantis brings the software and orchestration layer: cluster management, Kubernetes, enterprise cloud infrastructure, managed services, and the control plane needed to operate large-scale AI cloud environments.
Together, they start to look like a full-stack AI factory platform.
That is the strategic jump I think IREN is making.
And NVIDIA has its own reasons to want this.
If the bottleneck is shifting to infrastructure, NVIDIA cannot think only like a chip supplier. It needs places where its chips can become live compute at scale.
But more importantly, it needs non-hyperscaler, NVIDIA-aligned capacity.
That is what makes IREN strategically interesting.
IREN has no competing AI chip agenda.
It has power-backed capacity, large-scale sites, and execution speed.
That is what makes the 5GW partnership so important.
This is not just NVIDIA validating IREN.
It is also NVIDIA trying to secure and shape scarce deployment capacity for large-scale AI factories before that capacity becomes controlled by players who want competing silicon optionality, or are actively trying to reduce dependence on NVIDIA.
That is where the trio starts making strategic sense.
IREN has the sites and power.
NVIDIA has the AI factory blueprint and ecosystem.
Mirantis gives IREN the enterprise cloud operating layer needed to make the whole thing usable at scale.
If IREN can combine scarce power, fast buildout, NVIDIA-aligned architecture, managed cloud software, and enterprise-grade operations, it becomes one of the few companies capable of turning electrons into usable AI compute at industrial scale.
From $IREN news releases:
The partnership is intended to accelerate deployment of large-scale AI factories by combining NVIDIA’s DSX AI factory architecture with IREN’s expertise across power, land, data centers, GPU deployment and infrastructure operations.
Future deployments are expected to focus on IREN’s 2GW Sweetwater campus in Texas, which the companies expect to serve as a flagship deployment for NVIDIA’s DSX architecture.
Jensen Huang, Founder and CEO of NVIDIA, commented:
“AI factories are becoming foundational infrastructure for the global economy.”
“Deploying these systems at scale requires deep integration across the full stack — compute, networking, software, power and operations. IREN brings the scale and infrastructure expertise to help accelerate the buildout of next-generation AI infrastructure globally. Together, we are building for the age of AI.”
For all those claiming $IREN gave NVIDIA a “free call option”:
The 5-year option has value in theory. But in practice, that value becomes far less meaningful if both NVIDIA and $IREN believe $IREN clearing $70 within 5 years is highly likely.
NVIDIA sees the AI demand curve. They see the supply crunch. They see the data center bottleneck. They know who has powered land and who is just selling conference slides. They also just signed a 5-year, ~$3.4B AI Cloud contract with $IREN.
Don’t you think they believe $IREN can clear $70 over the next 5 years in this environment, given $IREN’s portfolio and capabilities?
I can assure you, the option is not some massive gift to NVIDIA. If anything, they would probably rather buy in at yesterday’s close price.
It is $IREN extracting value the most from this deal, not the other way around.
$IREN did not need NVIDIA’s cash. They said they had ~$2.6B of cash on hand. So instead of selling equity at ~$57, they pushed NVIDIA’s entry price to $70.
That spread alone is roughly ~$400M of potential value preserved versus issuing the same 30M shares around yesterday’s close.
And that is before counting everything else $IREN gets from the structure: NVIDIA as a customer, a 5-year ~$3.4B AI Cloud contract, validation from the most important company in AI, GPU roadmap credibility, financing credibility, more confidence from future counterparties, and a stronger position inside the AI infrastructure supply chain.
$IREN absolutely killed it last night. I’ve never been more bullish.
In fact, all my price targets for the next 5 years just went up significantly.
The company is firing on all cylinders and making all the right strategic moves.
Jacques B. retweeted
100 gigawatts under contract. 10 gigawatts of capacity left through 2030. Pricing up double digits. Competitor literally stopped taking orders. And they generated more free cash flow in 90 days than the prior 365. This market is the tightest it's been in decades and nobody's talking about it.
$IREN I think we’re getting close here.
The market is starting to connect the dots. It just needs one more datapoint from IREN before it rerates hard.
Based on how things have been evolving, I have high conviction the rerate will happen anytime between now and next earnings.
$IREN is about to explode. I feel all the pressure that’s built up over the past few months is about to be released and hit bears in the face.
I told you the other day that infrastructure growth will move much more slowly than chip output growth. Over time, that means the time to compute bottleneck shifts to infrastructure: AI ready, energized data centers. And when that happens, owners of energized AI data centers start capturing a larger share of the economic surplus across the AI value chain.
We are already seeing signs of that shift. The news is all around you.
I believe $IREN is the player best positioned for this today, because it solved for the physical bottlenecks years ago, when almost nobody saw this coming. I shared an interview nearly 6 months ago of Dan describing how this would unfold, and that interview is more than a year old now.
The value of every unsold MW $IREN still has today is materially higher than it was 6 months ago. And this trend will likely continue, because the infrastructure side of the equation is not getting solved anytime soon. It will actually get tighter from here.
$IREN still has ~4GW unsold (+pipeline). Every tranche will likely end up being monetized at a higher value than the one before it.
It will be obvious in hindsight.
Half of America's AI data centers planned for 2026 are delayed or cancelled. They're waiting on transformers. I build chemical plants. Transformer prices have tripled in the last four years. Lead times are 2 to 4 years. Each new plant we build competes with AI data centers for the same grid equipment. Every large power transformer in America runs on grain-oriented electrical steel. It's made by rolling iron and silicon together until their crystals align in one direction. No other alloy works at utility scale and only one US company makes it: Cleveland-Cliffs. The average large power transformer on the grid is 38 years old. Service life is 40. Amazon, Google, Meta, and Microsoft committed $650 billion to AI infrastructure this year. Nvidia's most expensive GPU is useless without a transformer.
I was too optimistic with my 2-year runway. Was more like 2 days 🤣
I think this is a much bigger deal than most investors realize. You can’t print data center capacity out of thin air. It will take a long time for this to be resolved (think years).
I think chipmakers and neoclouds with fake capacity will get hit badly.
$IREN should be one of the biggest winners here.
$NVDA GB300 Shipments Are Slowing as Data Center Infrastructure Lags
According to SemiAnalysis, shipment estimates for both the GB300 and VR NVL72 have been revised, with GB300 shipments now showing a slowdown
The issue is not demand, but infrastructure readiness. Some cloud service providers still do not have their data center infrastructure fully operational, which is causing a slight delay in overall shipments compared to previous expectations
Controversial prediction:
In the near future, say within the next 2 years, the AI industry’s growth projections will come down sharply because AI ready data center capacity will not scale fast enough to support them. There simply will not be enough infrastructure.
That mismatch will send pressure across the AI ecosystem. Chipmakers, much of the value chain around them, and companies that depend on access to affordable compute will get hit.
AI ready data center and infrastructure owners with real, deployable capacity will hold pricing power for a long time and be among the biggest winners.
$IREN When Dan says time to compute is everything, he is indirectly saying that chips alone are not enough. The last gate to compute is live AI-ready, energized infrastructure. Without that, chips do not become compute.
So at a high level, time to compute can be simplified as:
1.time to chips
2.time to AI-ready, energized datacenters
I think it is important to look at the AI value chain through that lens, because over time, 1 is likely to expand faster than 2, which means 2 is likely to become, or remain, the tighter bottleneck for much longer.
Why?
Because time to chips is mainly a manufacturing problem. It can improve through higher capex, better yields, added tool capacity, process optimization, and supply-chain reprioritization.
The datacenter side is different. It is an infrastructure problem.
It depends, among other things, on access to new suitable land, grid access, interconnection approvals, substations, transformers, switchgear, cooling systems, construction, and commissioning.
That is a heavier chain of dependencies. More importantly, many nodes in that chain are inherently slow. They are constrained by bureaucracy, permitting, coordination, and the physical world. Access to new suitable land is not solved in a couple of months. Neither are interconnection approvals, substations, major electrical equipment, construction, or commissioning. And unlike chips, those dependencies are local, physical, and often heavily regulated.
In Texas, transmission needed to support large new loads can take years to energize. And that is only one element of the AI-ready, energized datacenter side of the equation. If someone decides to build a datacenter from scratch today, getting the shells ready for chips can take years on top of that. That tells you a lot about the pace at which the infrastructure side can expand.
If compute demand keeps expanding faster than supply, the bottleneck will ultimately settle where expansion is hardest. Today, that increasingly appears to be AI-ready, energized datacenter capacity.
When Satya Nadella said Microsoft could end up with chips sitting in inventory because it did not have enough power-ready capacity, or enough “warm shells,” to plug them into, that was a clear signal that the time-to-compute bottleneck was shifting toward the infrastructure side of the equation.
And the slower that layer expands relative to demand, the more pricing power and economic surplus it is likely to capture from the value chain. The players who can supply live AI-ready capacity efficiently may end up controlling one of the most valuable choke points in the AI value chain.
Industry
• 50% of U.S. data centers expected for 2026 likely delayed or cancelled
• GPU rental pricing moving sharply higher
• Frontier model providers managing usage around peak compute windows
• AI use cases expanding rapidly and increasingly requiring more compute
• Microsoft said some GPUs are sitting in inventory because there is not enough powered, ready capacity to plug them in
• Supply-demand gap likely to worsen from here
$IREN
• Says time to compute is everything
• Projects confidence in its ability to deliver on schedule
• Recently ordered 50k B300 GPUs
• Has 4.5GW of total capacity, with more than 90% still uncontracted
• Expects that capacity to come online in the next 3-4 years (+more in the pipeline)
• Says customer demand is very strong and that demand is not an issue
• Shows Horizons 1 to 4 coming together rapidly
• Now saying more is underway
Sometimes connecting the dots is not that difficult.