The Cost of a Shovel Just Went Up
GPU rental forecasts revised higher again. Replacement costs rising 40 percent. The fastest revenue ramp in enterprise history just validated the demand thesis. Here what it means for your portfolio
Two weeks ago I wrote that GPU rental prices were the most honest signal in AI infrastructure. Since then, three things have happened that materially strengthen the thesis, and I want to be direct about what they mean for anyone positioned in this space. This piece is a follow-up, but it is designed to stand on its own. If you take one thing from this letter, let it be this: the cost of building new AI infrastructure is rising faster than the market appreciates, and that has immediate, actionable consequences for how you should think about the companies that already own it.
What This Means for Your Portfolio
I am front-loading the actionable framework because I know many of you read these letters to sharpen your own process. Here are the four things I would be watching right now.
First, track the replacement cost curve. The cost of building a next-generation GPU rack has moved from roughly $6.5 million in January to quotes in the low-to-mid sevens today. High-bandwidth memory costs are projected to rise 40 percent for 2027 shipments. When replacement costs rise, the economic useful life of existing hardware extends. Every data centre that is already built, already energised, and already generating revenue becomes more valuable, not because anything about the asset changed, but because the cost of replicating it just went up. Look for companies with large installed GPU fleets and near-term capacity additions. They are the direct beneficiaries of this dynamic.
Second, follow the rental data. One-year H100 contract prices have now been revised upward twice in consecutive model updates. The latest forecast calls for $2.75 per GPU hour by December 2026, up from $2.44 just weeks ago. Rental prices are a real-time clearing price for physical compute. They cannot be fabricated. When they rise, margins expand mechanically for every provider with capacity coming off contract. The market structure matters here: most GPU rental volume is transacted on contracts of one year or longer, not on-demand spot pricing. On-demand listings only adjust after the contract market has already moved. The contract data leads. And the contract data is saying that providers can now negotiate for higher prepayments, longer tenors, and more favourable start dates. Time is on their side. If rental prices plateau at current levels, the economics are still materially better than what most models assume. The only scenario that undermines the thesis is a meaningful decline in rental rates, and the supply chain data suggests that decline is unlikely before late 2027 at the earliest.
Third, distinguish between demand risk and engineering risk. The market has historically priced AI infrastructure as though demand might evaporate. But $625 billion in signed Microsoft backlog and $240 billion in Google backlog are not demand that can evaporate. It is committed capital. The real risks are execution risks: can providers build fast enough, can they secure memory and components, can they energise on schedule? These are engineering problems, and they favour companies that are already built, already powered, and already contracted. A farmer does not abandon his land because one season is dry. He assesses the aquifer, checks the soil, and plants again.
Fourth, watch the power pipeline. In a world where every hyperscaler is supply-constrained and new grid connections take four or more years, the companies that control energised power capacity hold a structural advantage that no amount of capital expenditure can quickly overcome. Power is the new moat. The chips will change, the models will evolve, the customers will rotate. But the power stays. If you find this analysis useful, I encourage you to subscribe to follow the thesis as it develops. The data will continue to update, and I intend to write every time the picture changes materially.
The Three Data Points That Changed This Week
Now let me walk through the evidence in detail, because each of these data points reinforces the others.
The first is the rental price revision. Industry data from the GPU contract market shows H100 one-year rental prices continuing to climb, with the latest forecasts revised upward for the second consecutive update. The new target of $2.75 per GPU hour by December 2026 represents a 7 percent increase from the prior model just weeks earlier. This is not a forecast built on assumptions about demand growth. It is constructed from contracted transaction data across over one hundred market participants, validated by actual deals. Two consecutive upward revisions in a matter of weeks. The direction of travel is unambiguous.
The second is the cost escalation across the entire hardware stack. High-bandwidth memory, the specialised memory that sits alongside every AI processor, is now projected to rise roughly 40 percent for 2027 deliveries. HBM represents a significant portion of the chip bill of materials, and the increase flows directly into GPU package costs, server pricing, and rack-level capital expenditure. Nvidia’s Rubin racks, which were being quoted at $6 to $6.5 million in January, are now drawing quotes in the low-to-mid sevens. For Rubin GPU packages specifically, this translates to roughly $7,000 in additional cost per unit. The blended increase is modestly lower than the headline 40 percent due to shifting supplier mix assumptions, but the direction is clear: the replacement cost of AI infrastructure is rising, not falling.
The third is what happened at Anthropic. The company’s annualised revenue crossed $30 billion in April. Up from $9 billion at the end of 2025. Up from $1 billion in January of last year. That trajectory represents one of the fastest revenue ramps in the history of enterprise software. Salesforce took approximately twenty years to reach the same figure. More than 1,000 enterprise customers now spend over $1 million annually, a number that doubled in under two months. Claude Code alone generated over $2.5 billion in run-rate revenue by February. Altimeter Capital’s Brad Gerstner said publicly that he would not be surprised if Anthropic exits 2026 at $80 to $100 billion in annualised revenue. I do not know if that number proves accurate. But I know what it implies about the demand for compute. Every dollar of AI revenue requires physical infrastructure underneath it. Servers, power, cooling, interconnection. And Anthropic is one company among many. OpenAI sits at $24 billion in run-rate revenue. The demand for the infrastructure that supports these businesses is not theoretical. It is contractual and it is accelerating.
None of these are forecasts. The rental revision is based on contracted transaction data. The HBM cost increase is based on supplier pricing negotiations. The Anthropic revenue figure was confirmed by Bloomberg. I am not building an argument on projections. I am building it on what has already happened.
The Replacement Cost Thesis
This is the dynamic that I believe has been most underappreciated. The conversation about GPU infrastructure has been dominated by one concern: that new supply will flood the market and compress margins. The data tells a different story. Not only are rental prices for existing hardware rising, but the cost to build the next generation of that hardware is rising in parallel.
When the replacement cost of an asset rises, the economic useful life of the existing installed base extends. An H100 that was deployed two years ago does not become less valuable when its replacement costs more. It becomes more valuable, because the hurdle rate for displacing it has just gone up. This is not different from a building that appreciates when construction costs rise. The asset you already own is repriced upward by the market’s inability to cheaply replicate it.
There is a second-order effect that deserves attention. When replacement costs rise, the useful life assumptions embedded in financial models should extend. A GPU that costs more to replace will be operated longer. A fleet that generates cash flow for six years instead of four changes the net present value calculation substantially. The industry data confirms this logic explicitly: rising rental rates improve return on invested capital by expanding margins on already-deployed fleets, while simultaneously extending the economic useful life of existing GPUs, meaning invested capital generates cash flows for longer before requiring reinvestment.
The memory shortage driving these cost increases is not a temporary disruption. SK Hynix, Samsung, and Micron have effectively sold out their HBM capacity through 2026. All three manufacturers are reallocating wafer capacity from commodity DRAM toward HBM, which consumes approximately three times the wafer space per gigabyte. This is a zero-sum reallocation. Every wafer dedicated to an HBM stack for an AI accelerator is a wafer denied to the DRAM modules used in laptops, smartphones, and conventional servers. Conventional DRAM contract prices are projected to rise 55 to 60 percent in early 2026, with server DRAM climbing more than 60 percent. The memory manufacturers have made their choice: they are building for AI margins, not commodity volumes. Relief is not expected until new mega-fabrication facilities reach volume production in late 2027 at the earliest. This is not a supply chain hiccup. It is a structural reallocation of the world’s memory manufacturing capacity toward AI, and it has consequences that will persist for years.
The Demand Validation
I want to spend a moment on the Anthropic number because it deserves careful attention. Roughly 80 percent of their revenue comes from enterprise customers. These are not casual subscriptions that churn when novelty fades. They are multi-year API contracts, cloud provider integrations, and signed enterprise commitments. Eight of the Fortune 10 are now Claude customers. Business subscriptions to Claude Code have quadrupled since the beginning of 2026.
Connect this back to the infrastructure thesis. Every token served requires a GPU. Every enterprise customer running agentic workflows requires sustained, high-concurrency compute. Anthropic has signed a 3.5-gigawatt compute deal with Google and Broadcom, with capacity expected to come online beginning in 2027. The fact that one of the fastest-growing companies in the world is locking in multi-gigawatt capacity two years in advance tells you everything about the supply-demand imbalance.
And Anthropic is one company. OpenAI disclosed $2 billion in monthly revenue. Microsoft’s Azure backlog stands at $625 billion. Google’s cloud backlog surged 55 percent in a single quarter. These are not forecasts. They are contracts.
I want to address the counterargument directly, because I take it seriously. The bear case says these revenue numbers are unsustainable, that AI spending will cool, that enterprises will pull back once the novelty fades. I understand the instinct. Rapid growth invites scepticism. But when 1,000 businesses are each spending over $1 million annually, when that number is doubling every two months, when the revenue is embedding itself into code repositories and business-critical workflows, you are not looking at novelty-driven demand. You are looking at enterprise adoption that compounds. The kind of demand that does not reverse easily.
The Power Ceiling Hardens
In my prior piece, I described power as the new moat. The latest industry modelling reinforces this by introducing behind-the-meter gas turbine options as an alternative to grid power. This is not an academic exercise. It reflects the fact that behind-the-meter strategies are gaining real traction in the United States because grid capacity simply cannot keep up with AI deployment timelines.
Grid electricity costs currently sit around $0.087 per kilowatt hour for data centre operators. Behind-the-meter gas turbine power can be generated at roughly $0.064 per kilowatt hour under base assumptions, but actual costs swing dramatically by region and season, from sub-$4 per MMBTU in oversupplied markets like ERCOT to over $12 in constrained Northeast winter conditions. The geopolitical dimension adds further uncertainty. Energy price volatility driven by the conflict in Iran creates planning risk for any operator dependent on gas-fired generation.
For an infrastructure provider with locked-in renewable power at $0.033 per kilowatt hour, this volatility is irrelevant. A company sitting on nearly 3 gigawatts of secured grid-connected power at that cost is operating at roughly 40 percent of the cost that a new entrant would face trying to self-generate with gas turbines. That is not a narrow advantage. It is a structural moat that deepens with every passing quarter, and it is insulated from a category of geopolitical and commodity risk that new entrants must absorb entirely.
The Instruments Are Confirming
In recent weeks, the market has begun to catch up. Several infrastructure providers have re-rated 30 to 35 percent as investors start to recognise the dynamics described above. The question is whether the re-rating is complete, and I do not believe it is. The rental price revisions, the replacement cost increases, and the demand validation from the leading AI labs all suggest that the fundamental picture has improved faster than the stocks have moved, even after the recent run. When rental prices are revised upward twice in three weeks and memory costs are forecast to rise 40 percent for the next generation of hardware, the financial models that drove the prior valuations need to be rebuilt. Most have not been.
I return to the image of the instruments in the cockpit. Two weeks ago, the market’s inner ear was telling it that AI infrastructure was tilting toward a bust while the instruments said otherwise. The instruments were right. The rental data confirmed it. The cost data confirmed it. The revenue data confirmed it. The market has begun to trust the gauges, but the altitude indicator is still reading higher than where the aircraft is flying.
The risks remain real, and I take them seriously. GPUs depreciate. Technology evolves rapidly. Execution on construction timelines is never guaranteed. Geopolitical disruptions to energy markets add another layer of unpredictability. These are engineering risks, and they deserve rigorous analysis. But they are not demand risks. And engineering risks favour companies with strong balance sheets, purpose-built infrastructure, physical assets that retain value beyond any single generation of silicon, and locked-in power costs that insulate them from commodity volatility. The companies that control energised capacity today are not just ahead on the deployment timeline. They are structurally insulated from the cost pressures that every new entrant must absorb.
The cost of a shovel just went up. The gold is still real. And the ground has not moved.
I do not know what the market will do next quarter. I never have. But I know what the rental rates are doing, I know what the replacement costs are doing, I know what the revenue of the largest AI companies is doing, and I know the difference between a thesis built on hope and a thesis built on arithmetic. Ours is built on arithmetic. That is enough.
Neel Khokhani
Founder and CEO, Epochal Corporation
@neel_epochal