Trends›Frontier training reaches $1T by 2030
Frontier training reaches $1T by 2030
Frontier model training scales by brute force (each generation costing ~10x the last), so gen-6 needs $100B clusters drawing ~10GW by 2028 and gen-7 reaches ~$1T cumulative capex by 2030, with power (not silicon) the binding constraint.
66 confidence, computed 23 Oct 2026
- Economic
- High weight
What rests on thismarkets it underwrites or threatens
Evidencethe schedule the figure draws
The evidence, newest first
- Jul 2026
against · observable Kimi K3 (Moonshot, 15 Jul 2026): 2.8T-param MoE, 16 of 896 experts per token, priced at Sonnet parity, claiming ~2.5x better scaling efficiency than K2 via Kimi Delta Attention (linear attention, fixed-size state replacing the growing KV cache) plus Gated… The Once and Future Kimi (State of the Future) - Jul 2026
against · claim OpenAI researcher argues scaling is entering a DATA-limited regime, not a compute-limited one: only ~300T tokens of useful public human text remain and RL's hard-task supply is drying up, so the binding constraint on capability shifts from cluster size to… A Stargate for Data — will depue (OpenAI) - Jun 2026
for · observable Chinese open-weight models now sit within 1-8 percentage points of Western frontier on reasoning benchmarks and achieve parity on coding at 5-30× lower cost per token. Chinese Open-Weight Models Are Nipping at the Heels of Western SOTA - Jun 2026
against · primary Four hyperscalers will pour roughly $725 billion into AI infrastructure in 2026—up 77% from ~$410 billion in 2025—with analysts projecting capex topping $1 trillion in 2027. Big Tech AI Capex 2026: Microsoft, Google, Meta, Amazon at $725 Billion - Jun 2026
for · primary Chinese open-weight models reached 61% of all tokens consumed on OpenRouter by May 2026, up from negligible share in late 2024. China's Open-Weight Takeover - May 2026
against · claim 2026 consensus shifting from single-axis pretraining toward inference/test-time compute and sparsity; analysts project inference ~75% of total AI compute by 2030. The one-axis era is over: AI's new scaling laws (synthesis) - Apr 2026
for · observable Meta raised 2026 capex to up to $145B and committed to building frontier 'superintelligence' models — a second well-capitalized actor pursuing brute-force scale. Meta Q1 2026 earnings call — superintelligence + up to $145B capex - Mar 2026
for · observable Huang projected ~$1T cumulative Blackwell+Vera Rubin orders through 2027 and framed inference compute as the next trillion-dollar driver. Huang bets ~$1T on inference / GTC 2026 (CNBC) - Mar 2026
neutral · observable SemiAnalysis argues the $5.6M DeepSeek figure excludes infrastructure, experimentation and failed runs (true cost far higher), tempering the 'scaling is dead' read. DeepSeek Debates: true training cost (SemiAnalysis) - Jan 2026
against · observable DeepSeek-class efficiency (671B-param MoE, 37B active, ~$5.6M reported compute) shows frontier-adjacent capability without nine-figure runs. DeepSeek cuts training costs with novel architecture (coverage) - Oct 2025
for · primary Five additional US Stargate sites (22 Oct 2025) bring planned capacity to ~7GW; program targets ~10GW under a ~$500B commitment. OpenAI/Oracle/SoftBank expand Stargate with five new sites - Jul 2025
for · observable Oracle committed to 4.5GW additional Stargate capacity (part of $300B+ over five years). Stargate advances with 4.5 GW Oracle partnership (OpenAI)
What would move itthe measures watched
- watchNamed gigawatt-cluster commitments and GW online (Stargate sites). GW announced vs GW actually energized; needs ~10GW single-purpose clusters by 2028
- watchDisclosed/estimated frontier training-run cost per generation. If each flagship costs ~5-10x the last, brute-force holds; frontier-parity at a fraction of cost weakens the $1T premise
- watchEfficiency breakthroughs at low cost (DeepSeek-class). Each credible 'frontier parity at a fraction of compute' result is evidence AGAINST needing $1T clusters
- watchTraining-vs-inference compute mix. If inference rises toward ~75% by 2030, the monolithic brute-force-training framing partially de-rates
- watchOpenRouter Chinese-vs-US share of production inference tokens (proxy for prediction 5). Chinese/open-weight share trending toward and past 50% = open-weight commoditising the closed frontier (prediction 5 resolving correct), softening the $1T-cluster slices; US/closed share holding = capex-determinism intact. Caveat OpenRouter skews to cost-sensitive indie/agent traffic, so it overstates enterprise migration.