Leading AI company reaches $40B compute costs, 6GW power, $200B capex
OPENBRAIN'S COMPUTE COSTS $40B 2026 ANNUAL · OPENBRAIN POWER REQUIREMENT 6GW PEAK POWER · CAPITAL EXPENDITURE $200B COST OF OWNERSHIP OF OPENBRAIN'S ACTIVE COMPUTE
At a glance
- Assessment: On Track
- Confidence: 55%
- Predicted timing: End of 2026
- Primary source: AI 2027, page 9, KEY METRICS 2026 dashboard (Late 2026: AI Takes Some Jobs)
What AI 2027 Predicted
The scenario’s KEY METRICS 2026 dashboard includes three infrastructure metrics for “OpenBrain” (the paper’s stand-in for a leading AI company). These describe the operational scale of the single dominant AI lab by end of 2026:
- $40B annual compute costs — the direct cost of running and maintaining their compute fleet
- 6GW peak power — the maximum power draw of all their datacenter infrastructure
- $200B capital expenditure — the total cost of ownership of their active compute hardware
These figures sit alongside other KEY METRICS (the separately tracked $45B revenue, $1T global capex, 38GW global power, and 2.5% US power share). While the global metrics are tracked as separate predictions, these company-specific figures represent the scenario’s view of how concentrated AI infrastructure becomes.
How We Track This
- Quarterly earnings and capex disclosures from Alphabet, Microsoft, Amazon, Meta (the hyperscalers hosting AI workloads)
- OpenAI infrastructure plans and spending (Stargate project, partnership with Microsoft/SoftBank)
- Anthropic infrastructure partnerships (Amazon/AWS, Google Cloud)
- Industry analyst estimates (SemiAnalysis, Futurum Group, CTVC)
- Datacenter power tracking databases and utility filings
- Individual datacenter project announcements with MW/GW capacity
Current Evidence
Compute Costs ($40B target)
No single AI company has publicly disclosed $40B in annual compute costs. Proxy estimates:
- OpenAI reportedly spends heavily on compute via Microsoft Azure but specific figures are not public. Revenue was $20B ARR at end of 2025, and costs are believed to exceed revenue substantially.
- Major hyperscalers’ total AI-related compute spending is in the $50-100B range combined, but this serves many customers, not a single company’s own research/deployment.
- The $40B figure implies a company spending roughly the equivalent of Anthropic’s + OpenAI’s entire combined compute budget on its own operations — this scale has not been reached.
- OpenAI projects $46B total compute costs for 2026 ($32B training + $14.1B inference) — exceeding the $40B threshold if accurate.
Power (6GW target)
- Individual modern AI datacenter campuses consume 500MW to 1GW
- No single AI company operates 6GW of peak AI-dedicated power as of early 2026
- Alphabet plans $175-185B capex in 2026, which could support multi-GW scale, but this serves all of Google’s operations, not AI R&D alone
- The CTVC database tracks 294 datacenter projects totaling 73.6GW planned capacity across all companies — but most are not yet operational
- Estimate: the largest single-company AI power footprint is likely in the 1-3GW range for leading hyperscalers, well short of 6GW
Capital Expenditure ($200B target)
The $200B capex figure for a single company is the most aggressive of the three targets:
- Amazon: ~$200B projected capex for 2026 (most but not all for datacenters; serves AWS broadly, not just AI)
- Alphabet: $175-185B capex for 2026
- Microsoft: tracking toward $120B+
- Meta: $115-135B
- Combined Big 5: $660-690B total in 2026
Individual company capex is approaching the $200B range, but this includes all infrastructure (not just AI compute ownership), and serves millions of customers — not the single-company-own-compute scenario AI 2027 describes.
Financing Structures
Axios reported in June 2026 that Apollo and Blackstone partnered with Broadcom on an AI infrastructure platform backed by an initial $35B loan to help Anthropic lease Google-developed chips through Fluidstack data centers. Axios says the platform is designed to enable more than 20 GW of compute capacity through 2028. This strengthens evidence that leading labs are using large private-credit and off-balance-sheet structures to reach tens of billions in compute financing. The reported 20 GW figure appears to describe broader platform capacity through 2028, not operational Anthropic power draw by the end of 2026.
Counterevidence & Limitations
- The “OpenBrain” figures describe a fictional company that is both the leading AI researcher and a massive infrastructure operator. No real-world analog exists — OpenAI relies on Microsoft’s infrastructure, Anthropic on AWS/Google Cloud. The prediction implicitly assumes vertical integration at a scale not yet seen.
- Hyperscaler capex is approaching $200B per company, but this serves entire cloud businesses, not a single AI research operation
- The 6GW figure is plausible for a hyperscaler’s total AI footprint by late 2026 but represents a much larger infrastructure than any current AI-focused operation
- Power grid constraints remain a binding limitation — many announced datacenter projects face 2-4 year interconnection queues
- Comparing AI-only costs vs. total company costs is inherently ambiguous; the source doesn’t fully distinguish between OpenBrain’s AI research compute and its inference/deployment compute
What Would Change Our Assessment
- Upgrade to On Track: A single company (or tightly coupled partnership like OpenAI+Microsoft) discloses or is credibly estimated to operate $20B+ in AI compute costs, 3GW+ AI power, and $100B+ AI-dedicated capex
- Upgrade to Confirmed: Figures approaching $40B/$6GW/$200B for a single entity’s AI operations
- Downgrade to Behind: By end of 2026, the largest single-company AI infrastructure remains below $15B compute / 2GW power / $80B capex
Update History
| Date | Update |
|---|---|
| 2026-06-15 | Axios reported a $35B Anthropic-linked compute-financing platform involving Apollo, Blackstone, Broadcom, and Fluidstack, designed to support more than 20 GW of compute capacity through 2028. This strengthens evidence for large compute-financing structures, while not proving Anthropic has reached the end-2026 $40B compute-cost, 6 GW power, or $200B active-compute thresholds. Confidence adjusted 0.50 → 0.55. |
| 2026-03 | OpenAI closes $122B funding at $852B valuation. Stargate program now encompasses ~10GW planned capacity. Assessment: $40B compute costs likely met (OpenAI projects $46B). 6GW in planned capacity exceeded but operational capacity harder to verify. $200B single-company capex approached by Amazon but serves entire AWS business. Status: Emerging → On Track for compute costs; power and capex remain Emerging. Confidence 0.40 → 0.50. |
| 2026-02 | Alphabet guides 2026 capex at $175-185B (nearly double $91.4B in 2025). Amazon guides $200B for 2026 (+53%). Big 5 combined: $660-690B. OpenAI revises compute spending to ~$600B by 2030 (down from $1.4T), projects $46B total compute costs for 2026 ($32B training + $14.1B inference) — exceeding the $40B threshold. Anthropic closes $30B Series G at $380B valuation. |
| 2026-01 | Meta guides 2026 capex at $115-135B (nearly double 2025). Stock surges 10.4%. |
| 2025-12 | Analysts project hyperscaler capex exceeding $600B in 2026 (+36% over 2025). SoftBank completes $22.5B second tranche to OpenAI. Michigan regulators approve 1.4GW for Stargate. Full-year 2025 capex: Amazon ~$131.8B, Alphabet $91.4B, Meta $72.2B. |
| 2025-11 | Sam Altman states OpenAI has ~$20B ARR and ~$1.4T in datacenter commitments over 8 years (~30GW capacity). Peak ambition before later revision. |
| 2025-10 | AMD and OpenAI announce 6GW GPU partnership (~$90B cumulative). OpenAI completes for-profit restructuring. Q3 capex: Alphabet $24B, Amazon $34.2B, Meta $19.4B, Microsoft $34.9B (+74% YoY). |
| 2025-09 | Anthropic closes $13B Series F at $183B valuation. First Stargate data center operational in Abilene, TX (200MW+). CoreWeave expands OpenAI deal by another $6.5B. Five new Stargate sites announced with combined ~7GW planned capacity. |
| 2025-07 | OpenAI and Oracle announce 4.5GW Stargate expansion exceeding $300B over 5 years. Alphabet raises 2025 capex guidance to $85B. Amazon reports Q2 capex of $31.4B, implying full-year ~$125B+. |
| 2025-05 | Q1 2025 earnings: Alphabet $17.2B, Meta ~$19.4B, Amazon $24.3B, Microsoft $21.4B quarterly capex. All confirm AI infrastructure acceleration. CoreWeave expands OpenAI agreement by $4B (cumulative ~$16B). |
| 2025-03 | OpenAI signs $11.9B five-year deal with CoreWeave. OpenAI closes $40B funding round at $300B valuation — largest private tech fundraise in history. |
| 2025-01 | Microsoft announces $80B capex for FY2025. Stargate Project announced at the White House — OpenAI, SoftBank, Oracle target $500B over 4 years with $100B immediate deployment. DeepSeek R1 released at claimed $5.6M training cost; Nvidia loses $589B in market cap. |