The AI industry is about to learn that money alone doesn't guarantee compute. In fact, the biggest bottleneck isn't GPU supply — it's timing.
Dylan Patel, CEO of Semi Analysis, has been tracking the capital expenditure plans of hyperscalers and AI labs with granular precision. His latest data reveals a staggering figure: the combined forecast for Amazon, Meta, Google, and Microsoft approaches $600 billion in capex this year alone. When you factor in the broader supply chain, total spending approaches $1 trillion.
The critical question isn't whether this compute exists — it's when it becomes available.
The Timeline Problem
A significant portion of that capex is committed to future capacity rather than immediate deployment. Google, for instance, has spent roughly $180 billion on turbine deposits for 2028 and 2029, data center construction extending into 2027, and power purchasing agreements that stretch further out. This isn't money deployed this year; it's money securing future slots in a queue that's growing faster than physical infrastructure can be built.
In America alone, approximately 20 gigawatts of incremental capacity will come online this year — a meaningful chunk driven by hyperscalers, but also including other players across the supply chain. The largest customers for that capacity are Anthropic and OpenAI. Currently, Anthropican is trying to scale to much larger than its present footprint.
The math gets uncomfortable quickly. If Anthropic adds $6 billion in revenue monthly — a conservative estimate given how aggressively AI adoption has grown — that's $60 billion of revenue across the next ten months. At current gross margins (at least as reported by Meta), that implies roughly $40 billion in compute spend to support that revenue. That $40 billion at $10 billion per gigawatt means Anthropic needs to add four gigawatts of inference capacity just to keep pace with revenue growth, assuming their research and training fleet stays flat.
The implication: Anthropic needs to reach well above five gigawatts by year's end — a target that's difficult but not impossible.
Why Conservative Strategies Backfired
Here's where the psychology gets interesting. Dario, who appeared on a previous podcast, was extremely conservative about compute commitments. He reasoned that if revenue flexed at a different rate, he didn't want to risk bankruptcy. He wanted responsible scaling.
But OpenAI took the opposite approach: signing aggressive deals regardless of immediate financial constraints.
The result? By year's end, OpenAI will have substantially more compute available than Anthropic — somewhere between two and three gigawatts more, based on current trajectory. This isn't just about raw capacity; it's about locking in favorable pricing before supply runs short.
When you sign contracts early, you negotiate from strength. When you wait until you're desperate, you pay premiums — whether through revenue-sharing arrangements with cloud providers like Amazon (Bedrock), Google (Vertex), or Microsoft (Foundry), or through last-minute spot pricing that carries a 50% markup over baseline costs.
The conservative approach worked beautifully as long as compute was abundant. Now it's scarce, and those who bet on scarcity are scrambling to acquire capacity from providers they previously wouldn't have touched: NeoClouds with lower-quality infrastructure, companies like SoftBank Energy who've never built data centers but are building them now for OpenAI, Oracle, CoreWeave, and others.
The Funding Paradox
OpenAI announced $110 billion in raises. Anthropic announced $30 billion. These figures sound massive until you realize what they have to cover: the yearly price of a one-gigawatt data center is approximately $1-3 billion. OpenAI alone will need roughly four gigawatts this year — which their raises can comfortably cover, even before considering this year's projected revenue.
But here's what's telling: Anthropic's conservative strategy means they're currently constrained. There simply aren't enough incremental buyers of compute to go around because Anthropic hit the capabilities tier first where their revenues exist. Having the best model is an extremely depreciating asset — three months later, you might not have the best model. But having capacity locked in matters enormously.
The Depreciation Question
One financial implication that's been strangely absent from mainstream analysis: what happens to GPU depreciation cycles?
Critics like Michael Bur have argued that GPUs should be depreciated over four or five years. But the reality is more extreme. With Hopper (H100) infrastructure, and especially if AI really takes off in coming years, the actual useful life of these chips extends far beyond typical accounting assumptions.
An H100 costs roughly $140 an hour to deploy at volume across a five-year depreciation schedule. If you sign deals at $2 per hour for those five years, your gross margin is approximately 35%. If you sign at $1.90, it's around 35% as well. By the fifth year, the GPU essentially falls off — it's dead in terms of useful compute life.
This implies that the depreciation cycle might actually be longer than five years, making reported amortized capex appear more favorable for cloud builders than actual economics would suggest.
Bottom Line
The core insight here isn't about hardware availability or supply chain constraints. It's about capital timing and strategic positioning. The labs that signed aggressive deals early have locked in compute at better prices and with better providers. Those that waited to be responsible are now paying premiums they couldn't have anticipated — both in dollars and in capability lags.
For anyone building AI infrastructure, the lesson is clear: there is no ethical way to scale compute conservatively when demand outpaces supply this severely. The bottleneck isn't GPUs. It's access to capital when you need it most.