One row per AI lab. Purpose: a comparable battery across…: How many gigawatts of dedicated AI compute capacity will…

Part of: One row per AI lab. Purpose: a comparable battery across new AGI-style labs with multibillion-dollar war chests, to gauge whether any could one day rival OpenAI or Anthropic in the race for transformative AI. Anchors for scale: OpenAI and Anthropic each operate or have contracted multiple gigawatts of compute (OpenAI's Stargate program targets 7+ GW with roughly 1 GW live in mid-2026; Anthropic has contracted up to 1 million Google TPUs plus multi-gigawatt expansions from 2027). Frontier training runs cost billions of dollars per model generation. Question per row: how many gigawatts of dedicated AI compute capacity will the named lab control, owned or contracted for its exclusive use and scheduled to be live, as of December 31, 2028? Resolve from credible reporting, company or partner announcements, and filings; convert reported chip counts to gigawatts using standard all-in power estimates (as used by Epoch AI or SemiAnalysis); include capacity under signed contract and live by the date; exclude options, letters of intent, and revoked deals. — view all 6 rows

How many gigawatts of dedicated AI compute capacity will Thinking Machines Lab control, owned or contracted for its exclusive use, as of December 31, 2028?✓#2

gw_capacity0.75GW

p10
0.2
p25
0.4
p50
0.75
p75
1.15
p90
1.7

rationale:

The Anchor Commitments The primary driver of this forecast is Thinking Machines Lab's (TML) March 2026 strategic partnership with Nvidia to deploy "at least one gigawatt" of Vera Rubin systems 2 sources. Deployments are targeted to begin in early 2027 2 sources. Alongside this, TML signed a "single-digit billions" expansion with Google Cloud in April 2026 for Nvidia GB300 NVL72 capacity aimed at reinforcement learning 3 sources. Based on multi-year rental pricing and all-in power estimates for GB300s, this Google Cloud commitment maps to roughly 50–200 MW of equivalent capacity 2 sources, though its non-exclusive nature raises questions about how much will qualify as "dedicated" 2 sources. Additionally, TML placed a smaller $471.7M order for 5,000 B300 GPUs (~0.01 GW) in May 2026 theinvestorchannel.substack.com.

Financial and Physical Ramp Constraints While the headline 1 GW figure is substantial, capacity must be scheduled to be live by December 31, 2028, to count. A 1 GW data center build costs approximately $50–60 billion 2 sources. TML's known funding is a fraction of this: a $2B seed round at a $12B valuation 2 sources, plus an undisclosed Nvidia investment 2 sources. Attempts to raise a ~$50B-valuation mega-round stalled by early 2026 amidst co-founder departures 3 sources. Furthermore, a multi-year procurement starting in early 2027 naturally implies a staggered ramp into 2029 and beyond, not a fully energized gigawatt by the end of 2028. Strikingly, no specific site, hosting partner, or power contract has been publicly identified for the Vera Rubin campus datacenterdynamics.com.

Distribution and Key Uncertainties Given these constraints, the median estimate (~0.74 GW) assumes a meaningful but incomplete Vera Rubin ramp—delivering a few hundred megawatts of live capacity by late 2028—plus roughly 0.1–0.2 GW of countable Google Cloud or neocloud capacity. We discount the headline 1 GW commitment due to the high likelihood that delivery schedules stretch well into 2029.

The downside tail (p10 of ~0.18 GW, p25 of ~0.39 GW) reflects scenarios where execution risks materialize: funding shortfalls prevent substantial deployments, data center build-outs are delayed, or a strict reading of the resolution criteria excludes the shared Google Cloud resources. Conversely, the upper tail (p75 of 1.15 GW, p90 of 1.83 GW) accommodates the possibility of an aggressive, Nvidia-backstopped mega-build where the full 1 GW is energized on schedule, supplemented by follow-on hyperscaler agreements, or a lenient resolution interpretation that credits the entirety of the signed "at least 1 GW" contract regardless of phasing.

Set against similar highly funded, vendor-backed peers, the distribution was marginally adjusted to reflect the high probability that a full 1 GW deployment schedule will stretch into 2029 due to physical constraints.