The Compute Crisis: Why Google and Anthropic Are Paying SpaceX a Premium to Rent GPUs

An Industry-Wide Compute Famine

Microsoft, Google, Anthropic — three titans of the AI world — are all grappling with the same thorny problem: a severe shortage of compute power.

Microsoft's CEO stated plainly that despite strong Q1 growth, the company expects to remain capacity-constrained throughout the first half of the fiscal year, even as new data center capacity continues to come online. In plain English: it's not that we can't make more money — we just don't have enough compute to make more money. Google CEO Sundar Pichai echoed the sentiment, saying the company remains supply-constrained even while continuously expanding capacity — and this is a company that manufactures its own TPU chips.

What happened next was even more shocking: Anthropic began paying SpaceX $1 billion per month to rent idle compute capacity. Shortly after, Google followed suit, purchasing compute from SpaceX at $920 million per month. The irony is rich — Anthropic had previously banned xAI and SpaceX from using its models over distillation concerns. Now, it has no choice but to bow to this very "adversary."

From Sand to GPU: A Full Breakdown of Supply Chain Bottlenecks

To understand the severity of this compute crisis, we need to dissect the entire GPU manufacturing supply chain. Every link in the chain is a potential bottleneck, and unfortunately, nearly every one of them is currently choking.

TSMC: The Absolute Core of Semiconductor Manufacturing

TSMC is the undisputed center of global semiconductor manufacturing. From Apple to NVIDIA, from AMD to Intel, every major chip company relies on TSMC to fabricate their chips. TSMC doesn't design chips — you hand over the blueprints, and they build the silicon for you.

Apple, thanks to its historically deep partnership with TSMC, has locked in substantial long-term capacity allocations and is relatively "safe" at this layer for now. But NVIDIA's demand for TSMC capacity is exploding, with requests for increased manufacturing share every few months. Despite this, TSMC's annual revenue growth is only about 40% — not because demand is insufficient, but because production capacity simply cannot keep up with the explosion in demand.

More critically, TSMC estimates that adding new manufacturing capacity requires an 8 to 10 year construction cycle. This means today's capacity shortage has virtually no short-term solution.

High Bandwidth Memory (HBM): Three Companies Control the World's Lifeline

Manufacturing GPUs requires more than just TSMC's silicon wafers — it also requires High Bandwidth Memory (HBM). Only three companies in the world can produce HBM: SK Hynix, Samsung, and Micron.

These three companies previously split their capacity between consumer devices (phones, SSDs, RAM sticks) and data centers. But AI's appetite for memory is simply insatiable — even smaller models like DeepSeek V4 Flash require over 100GB of VRAM — forcing these manufacturers to massively shift capacity from the consumer side to the data center side.

A landmark event: Crucial, Micron's consumer memory brand, has ceased operations. Micron decided to shut down Crucial's consumer business and redirect all resources toward higher-margin AI chip memory. Consumer memory prices have skyrocketed to jaw-dropping levels.

Hard Drives and Power: The Overlooked Hidden Bottlenecks

The compute crisis extends far beyond GPUs and memory. Hard drive manufacturer Western Digital's capacity is sold out through all of 2026. For perspective: refurbished 16TB drives previously sold for around $170, but the same spec now goes for over $360 — more than double.

Power supply is another severely underestimated bottleneck. U.S. commercial electricity demand is projected to surpass residential demand for the first time in 2027 — historically unprecedented. Compared to China, the U.S. lags far behind in grid expansion: the amount of new power generation China has added annually since 2016 exceeds the total increase the U.S. has achieved since the 1990s.

These major compute companies have already begun investing in power infrastructure themselves. Microsoft is pushing nuclear energy projects, while Musk is leveraging Tesla's battery manufacturing capabilities to provide backup power for xAI's data centers, even deploying gas generators — rather ironic for the founder of an electric vehicle company.

Why SpaceX Became the Biggest Winner of the Compute Crisis

Understanding why SpaceX came out on top in this crisis requires going back to a bet Musk made over a year ago.

Musk bet on the right trend, but the wrong use case. He was convinced that compute would become the biggest bottleneck, so he massively over-purchased GPUs for xAI and Grok, building a supercomputing cluster called Colossus. The initial construction cost was roughly $3–4 billion. However, Grok underperformed, top researchers at xAI departed one after another, and much of that compute sat underutilized.

But this very "mistake" turned into a goldmine. When the entire industry was desperate for compute, SpaceX was sitting on a massive stockpile of idle GPUs and began renting them out at $1 billion per month. In just four months, the entire construction cost of Colossus could be fully recouped.

Google pays nearly $12 billion annually for this — roughly 3% of its annual revenue. That figure is on par with what Google pays Apple for default search engine placement — one of the tech industry's most famous "protection fees."

Strategic Wins and Losses Across AI Companies

This compute crisis has exposed massive differences in strategic judgment across companies:

• OpenAI: Bet on compute and Scaling Laws years ago, stockpiling as much capacity as possible. This is why OpenAI's API rate limits are far more generous than Anthropic's — they started positioning well before the crisis hit.
• Anthropic: Exercised financial prudence and was relatively conservative in compute procurement last year. The result: compute that was available to buy last year is no longer available this year, forcing them to rent from SpaceX at a premium.
• Google: Assumed its in-house TPUs could bypass the bottleneck, but it turned out that custom chip production capacity fell far short of needs, ultimately requiring payments to SpaceX.
• NVIDIA: Regardless of how other bottlenecks shift, as long as demand exists, NVIDIA is the biggest structural winner. Its problem isn't selling chips — it's making them fast enough.

What the Compute Shortage Means for Everyday Users

The impact of this crisis is spreading from the enterprise side to the consumer side. Prices for RAM, hard drives, and SSDs are all climbing because manufacturers are redirecting capacity from consumer products to higher-margin data center products.

A practical piece of advice: if you've been waiting for hardware prices to drop before building a PC or upgrading storage, prices are unlikely to come down in the near term. Demand is too extreme, and every link in the supply chain needs years to scale up.

The deeper implication is that our phones and personal devices may not continue getting faster and more powerful at the pace we've been accustomed to. Instead, an increasing share of computation will shift to the cloud, concentrated in the hands of a few major players. Compute is moving from distributed to centralized, and this will profoundly reshape the entire tech industry landscape.

The Essence of the Compute Crisis: Systemic Bottlenecks Can't Be Solved One at a Time

At its core, this compute crisis is a systemic problem where all bottlenecks must be resolved simultaneously. Scaling TSMC capacity 10x is useless if HBM can't keep up. Scaling HBM 10x is useless if hard drives and power can't keep up. A shortfall in any single link becomes the ceiling for the entire system. And right now, nearly every link is a shortfall.

For the AI industry as a whole, compute is no longer just a technical issue — it's the strategic resource that determines who wins the next round of competition. Whoever locked in compute early holds the initiative for the future.