One row per AI lab; part of a comparable battery across…: When will Thinking Machines Lab first have a model at the…

Part of: One row per AI lab; part of a comparable battery across new AGI-style labs with multibillion-dollar war chests, gauging whether any could rival OpenAI or Anthropic. Question per row: when will the named lab first have a frontier model? Resolution: the earliest date at which either (a) a model from the lab appears in the top 5 of the Artificial Analysis Intelligence Index (or its successor flagship overall-quality leaderboard) at any snapshot, or (b) for unreleased or unlisted systems, credible reporting plus at least one independent expert assessment establishes that the lab has trained a model broadly regarded as frontier-class, comparable to the contemporaneous top models from OpenAI, Anthropic, or Google (the way Anthropic's unreleased Mythos-class model inside Project Glasswing was publicly established before launch). If the lab is acquired or wound down first, or the event seems unlikely for many years, express that through late percentiles. Anchors: the current index top-5 is occupied by Anthropic and OpenAI variants; frontier generations currently require training runs costing billions of dollars and multi-hundred-megawatt compute. — view all 6 rows

When will Thinking Machines Lab first have a model at the AI frontier, per the resolution criteria?✓#2

frontier_dateDec 1, 2030

10% by
Feb '28
25% by
Jan '29
median
Dec '30
75% by
Jan '35
90% by
Jan '43

rationale:

As of August 2026, Thinking Machines Lab (TML) has shipped foundation models but remains far from the AI frontier. Its strongest release, Inkling (a 975B-parameter MoE), recently scored 41 on the Artificial Analysis Intelligence Index artificialanalysis.ai. This sits roughly 18 points below the top-5 cutoff of 59, a tier currently saturated by Anthropic and OpenAI reasoning-effort variants artificialanalysis.ai. Because cracking the top 5 requires displacing these dominant players' variants, TML's most plausible path to resolving the question is through credible reporting and expert consensus that an unreleased internal model is frontier-class. However, TML itself acknowledges that Inkling is not the strongest overall model available today 2 sources.

The most significant upside catalyst for TML is its compute pipeline, which dictates the earliest possible timeline for a frontier breakthrough. In March 2026, the lab secured a multiyear partnership with Nvidia for at least 1 GW of next-generation Vera Rubin systems, explicitly targeted for deployment in early 2027 44 sources. Supplemented by an April 2026 Google Cloud deal for GB300-powered infrastructure techcrunch.com, TML will soon have the hardware capacity to train at the frontier. Factoring in physical deployment, pre-training, and post-training cycles, the earliest a genuinely frontier-scale artifact could emerge is late 2027 to early 2028. A highly successful, rapid run on this new hardware drives the early 2028 to early 2029 estimates.

Despite this strong compute access, TML faces severe headwinds that threaten its ability to match the trillion-dollar incumbents. Capital is a primary constraint: the company remains at its $12B July-2025 valuation after a $50B mega-round collapsed in early 2026 3 sources, yet it will require tens of billions of dollars to finance its hardware commitments reuters.com. Furthermore, the lab has suffered significant talent attrition, losing roughly a third of its founding team to competitors, including Barret Zoph, Luke Metz, Andrew Tulloch, and Lilian Weng 3 sources. Strategically, TML is deliberately betting on customization, interaction models, and open weights rather than raw leaderboard supremacy 2 sources—a niche that could prove highly successful commercially without ever satisfying the frontier resolution criteria.

Consequently, there is a substantial probability that this event never occurs. TML could be acquired (Meta reportedly floated a ~$1B offer previously, and hyperscalers remain likely buyers), could be wound down if it fails to finance its immense compute obligations, or could permanently persist as a profitable infrastructure layer sitting comfortably below the frontier 2 sources. This high non-occurrence risk pushes the median estimate out to late 2030, reflecting a blend of multi-generation catch-up scenarios and failure modes. The extremely late tail, spanning 2035 to 2043, effectively encodes these long-term acquisition, wind-down, and permanent second-tier outcomes.