Dylan Patel on space data centers, the TSMC bottleneck returning in 2027, and why OpenAI is fine

Summary

Dylan Patel, founder of SemiAnalysis, describes a semiconductor-constrained future arriving in 2027, argues that space-based data centers remain years away from meaningful scale, and dismisses Oracle's panicked messaging about OpenAI's financial stability.

The TSMC bottleneck returns in 2027

The energy debate obscures the real constraint. Power generation can scale. The US has enough turbine manufacturing capacity (Cummins alone makes a million diesel engines annually) and regulatory obstacles are solvable. Semiconductor fabs cannot scale the same way. In 2024–2025, the industry swung from chip shortages to power constraints. By 2027, the binding constraint shifts back to semiconductors.

TSMC has no slack capacity the way energy vendors do. You cannot call a broker and buy a three-nanometer fab at 2x markup the way you can with a power turbine. Memory makers have barely expanded since 2022. Building a new fab is the most complex building humans make, with clean rooms that circulate air every 1.5 seconds to maintain parts-per-billion particle counts. Even if makers wanted to double capacity today, construction takes years.

Space data centers: one order of magnitude off

SpaceX's Starship launch costs are falling fast enough to make orbital compute viable eventually. Heat dissipation and radiator engineering are solvable. The real problem is reliability and repair logistics.

Chips in space fail at higher rates than on Earth. Hopper GPUs see 10–15% RMA rates in the first two weeks after deployment, and Blackwell starts higher. Large chips fail more often. If bit error rates stay constant, a chip twice as fast fails twice as often. One 10x larger fails 10x more. You cannot send a technician to replace a failed GPU in orbit.

Patel sees a bet between xAI and Anthropic's heads of compute on whether 1% of global data center capacity will be in space by 2028. He takes the under, though the math is not wild—roughly 150 Starship launches to reach a gigawatt in space. The constraint is not launch cost or power (solar panels and launch economics favor orbit), but rather the engineering burden of dealing with unreliable hardware at scale, smaller clusters that cannot fully interconnect, and no service technicians.

Currently, Starlink v2 satellites draw about 25 kilowatts per ton for launch. If Starlink v3 doubles that to 50, and a future compute variant reaches 100 kilowatts per ton, the math works. That requires only 150 or so Starship launches. The timeline might not be 2028, but 2029 is plausible.

Chip diversification across the stack

Nvidia's Grok acquisition signals a strategic shift. For years, Nvidia pitched one GPU that did everything. Now it is launching CPX (a chip optimized for pre-fill and prompt processing), keeping the standard GPU line, and absorbing Grok chips. Each targets different niches. Patel reads this as Nvidia saying it does not really know exactly where AI is going. The company is hedging across the Pareto curve rather than betting on a single architecture.

Google diverged its TPU roadmap similarly. Broadcom and MediaTek now fab different TPUs at TSMC for different workloads. Google has a third TPU project underway. Everyone at sufficient scale is proliferating along the curve of high-flop low-memory variants, fast-on-chip-memory-only variants, 3D-stacked-memory variants, and general-purpose middle ground.

Cross-data-center training got easier

Google's regional clusters (data centers 40 miles apart in Nebraska, Iowa, and Ohio, with new builds in Oklahoma and Texas) solved some pre-training scaling challenges. But the bigger shift is the move from pre-training to reinforcement learning. RL training spends most compute generating data through forward passes, then sends only verified final tokens back for actual training. This requires far less synchronization overhead than pre-training weight updates.

As systems become more agentic, you may not need to send entire weights at all, just tokens relevant to the rollout. Synchronization drops from every 10–20 seconds to every few minutes. Multi-data-center training is completely reasonable now. Anthropic does this, running inference on one set of chips and training on another.

OpenAI and Cerberus: capacity, not crisis

OpenAI's Cerberus deal for 750 megawatts is real infrastructure and limited in scope. Latency does not drive long-horizon inference workloads; inference speed does. An agent running for 30 minutes versus 10 minutes versus 5 minutes is the value lever, not where the compute lives geographically. Some users are price-insensitive enough to pay 10x for 10x faster completion.

Cerberus can serve that edge case. But OpenAI must solve the product problem: how do you expose different latency tiers without adding toggles. Today, Codex has six models plus another button to pick from. OpenAI's marketing is chaotic. Anthropic has similar naming chaos (Claude, Claude Code, plus Claude Code for non-Claude-Code tasks).

Oracle's frantic posts assuring the market that OpenAI can meet its commitments amount to terrible comms. The lion should not concern itself with sheep. Nvidia ran the same playbook during TPU hype last year. Both companies should hire better PR.

China policy and the value-capture question

Patel frames the chip-export debate as a game-theory problem. Push China too hard into a corner and it escalates—more military action, deeper control of Africa and Latin America, or direct takeover of Taiwan. But allow China full access to AI and you accelerate its military integration of the technology.

Dario Amodei's view is to not sell chips or API access. Patel is more sympathetic to Ben Thompson's position: sell API access, not equipment. The logic is economic. Sell China tens of billions in lithography equipment and they capture hundreds of billions in chip and AI value. Sell them API access at per-token costs and you capture more of the value chain. China has historically refused dependency on American ecosystems (Windows, Visa) and built alternatives (bootleg OS, Alipay and WeChat Pay) that often outperform the originals. The question is where to draw the line so China must stay on the service layer rather than own the supply chain.

Anthropicb does not sell API access to China directly. Traffic routes through Korea and other intermediaries.

Jensen Huang's two modes

Patel observes that Jensen Huang operates in two distinct modes. In meetings, he is a business killer, deeply versed in design, manufacturing, energy, power, data center logistics, and supply contracts across the entire stack. Outside, he is the PR Jensen, good at stage presence and hype. Walking out of a negotiation meeting straight into street interviews with paparazzi creates a schizophrenic image: killing on the inside, then walking into the clip-farming circus. Nvidia should invest in better communications.