Thinking Machines Lab
Venture-funded, compute-intensive AI research lab monetizing via developer infrastructure (Tinker, a managed LoRA fine-tuning API for open-weight models, with published per-token pricing but undisclosed revenue) plus a planned move up the stack to proprietary multimodal 'interaction models'
Only the Jul 2025 seed is a completed priced round. The Feb 2025 ~$9B figure was reported pre-close talk, the Nov 2025 ~$50B was an attempted markup that collapsed by Jan 2026, and the Mar 2026 NVIDIA investment has no disclosed valuation - the trail therefore has one hard mark ($12B) bracketed by reported negotiation levels.
Earnings, margins, COGS & capex
Private and opaque. The only hard financial fact is the $2B seed at $12B post (Jul 2025, led by Andreessen Horowitz; NVIDIA, AMD, Accel, ServiceNow, Cisco, Jane Street, and Albania's government with $10M participated). Revenue, burn, and margins are not disclosed. The Nov 2025 attempt to raise at a reported ~$50B (some reports up to $55-60B) collapsed by Jan 2026 because prospective backers declined the price without a more substantial product record - the most informative external mark on the business. NVIDIA subsequently made an undisclosed strategic investment (Mar 2026). Compute commitments (1+ GW Vera Rubin, reported at tens of billions over the multi-year term; multi-billion Google Cloud deal) imply burn far ahead of the $2B raised, making another financing event structurally necessary.
Revenue trend
Margins
unknown
unknown
COGS structure
Dominated by GPU compute: NVIDIA Vera Rubin capacity (at least 1 GW multi-year commitment, announced Mar 2026, deploying from early 2027) and Google Cloud AI Hypercomputer infrastructure including early NVIDIA GB300 NVL72 access (Apr 2026, reported single-digit billions of dollars). Second-largest cost is elite research talent (~183 employees per Tracxn as of May 2026, heavily ex-OpenAI, at frontier-lab compensation).
Capex
As a private lab it rents/commits rather than builds datacenters, but the committed-compute intensity is extreme relative to stage: the FT reported the NVIDIA chip-supply arrangement alone at tens of billions of dollars, plus the multi-billion Google Cloud deal - effectively capex-like forward obligations against a $2B raise that require either revenue traction or a successful new round.
Latest earnings
none; the key stated milestone is taking interaction models from research preview to a broader release later in 2026
Growth drivers
- Tinker GA expansion (Dec 2025) — waitlist lifted, added trillion-parameter Kimi K2 Thinking, vision-language models (Qwen3-VL-30B/235B), large MoE models (e.g. Qwen3-235B-A22B), and OpenAI-API-compatible inference (beta)
- Interaction models (TML-Interaction-Small, announced May 2026) — full-duplex multimodal AI (276B-parameter MoE, ~12B active) claiming sub-0.4s turn-taking latency on FD-bench, positioned as a new category vs. sequential chatbots; limited research preview then broader release planned for later 2026
- Enterprise demand for fine-tuning open-weight models as a cheaper, controllable alternative to closed frontier APIs
- NVIDIA strategic investment + compute partnership giving preferential access to next-gen Vera Rubin capacity
- Research-community flywheel — $100K + Tinker-credit grants for human-AI interactivity research (announced by John Schulman, Jun 2026)
Bull & bear
The team is the trade: a deep ex-OpenAI research bench, $2B of runway, locked-in next-gen NVIDIA and Google Cloud compute, and a genuinely novel product thesis (interaction models) that - if the late-2026 launch lands - re-rates the company back toward the ~$50B mark investors briefly entertained.
- Schulman/Murati-grade research talent has repeatedly been the leading indicator of frontier breakthroughs; Schulman and Weng remain, and the lab kept shipping through the Jan 2026 turmoil
- Tinker is real, shipped, GA with published pricing, and well-reviewed by researchers - a distribution wedge into every org customizing open-weight models, a structurally growing segment
- TML-Interaction-Small's claimed sub-0.4s full-duplex turn-taking latency (vs. 0.57s Gemini-3.1-flash-live and 1.18s GPT-realtime-2.0 on first-party FD-bench) is a demonstrable technical lead in a category with obvious consumer and enterprise pull
- NVIDIA's investment plus the 1 GW Vera Rubin allocation and the Google Cloud GB300 deal solve the scarcest input (compute) on preferential terms - the most supply-constrained player in AI chose to fund and supply them
- Entry at the $12B mark is a ~4x discount to where sophisticated investors were negotiating in Nov 2025; the failed round was a price dispute, not a diligence scandal
A lab with no disclosed revenue burning frontier-scale compute, whose growth round already failed once, whose co-founders left (one fired), and whose two products face immediate commoditization - the $12B seed mark may still overstate what a strategic or down-round buyer would pay.
- Zero disclosed revenue at month ~17; Tinker publishes pricing but no customers or usage, and fine-tuning APIs are a commodity fought over by Together AI, Fireworks, Databricks, and the hyperscalers
- The market already voted: ~$50B talks collapsed by Jan 2026 because backers declined the valuation without a more substantial product record
- Co-founder departures (Tulloch to Meta; Zoph fired then rehired by OpenAI alongside Metz and Schoenholz) undermine the single thesis - the team - that justified the seed price
- Compute commitments reported at tens of billions against a fixed $2B raise create a financing treadmill; another failed round forces down-round, structure-heavy terms, or an acqui-hire outcome
- Interaction models are an announced research preview with first-party-only benchmarks, not revenue; OpenAI's realtime voice stack and Google's Gemini Live attack the same UX with billions of users of distribution
- Open-weight commoditization cuts both ways: if Chinese and Meta open models keep improving, value accrues to compute and distribution owners, not to a mid-scale post-training specialist
What it is worth
Last priced round, sanity-checked against the failed markup attempt and comparable frontier-lab marks; no revenue multiple is computable (revenue undisclosed)
$3-6B
further senior departures or another failed raise force a structured down-round or strategic/acqui-hire outcome; tens-of-billions compute commitments become liabilities rather than assets
$10-15B
holds around the seed mark; team and locked-in compute access justify roughly the last priced round while revenue proof remains absent
$40-60B
interaction models ship broadly in late 2026 with clear differentiation, Tinker shows disclosed commercial traction, and the stalled round re-opens near the previously discussed range
Only real mark: $12B post-money (seed, Jul 2025, $2B raised - largest seed round in history per Crunchbase). Nov 2025 talks at a reported ~$50B (up to $55-60B per some reports) collapsed by Jan 2026 - an explicit market rejection of the markup absent product traction. NVIDIA's Mar 2026 strategic investment is undisclosed in size and price. Any current value is a bet on team + compute access + the late-2026 interaction-model launch, not on financials, which do not publicly exist.
SWOT
Strengths
- Dense ex-OpenAI founding talent — Mira Murati (ex-OpenAI CTO), John Schulman (OpenAI co-founder, RLHF pioneer) as chief scientist, Lilian Weng - though thinned by departures
- $2B seed - largest in history - provides multi-year runway even at frontier-lab burn
- Strategic backing from NVIDIA (investor + 1 GW Vera Rubin commitment) and Google Cloud (GB300 NVL72 early access) secures scarce next-gen compute
- Shipped real products fast — Tinker (8 months post-founding, GA at month 10), respected open research (LoRA post-training, deterministic-inference/output-consistency work), Tinker Cookbook OSS credibility with researchers
Weaknesses
- Co-founder attrition — Andrew Tulloch to Meta (Oct 2025); Barret Zoph (fired amid an alleged confidential-information dispute) and Luke Metz returned to OpenAI with researcher Sam Schoenholz (Jan 2026); Fortune reported a wider wave of defections
- Failed markup raise — ~$50B talks (Nov 2025) collapsed by Jan 2026 - prospective backers declined the valuation without a more substantial product record; the $12B seed remains the only priced mark
- No disclosed revenue ~17 months after founding — Tinker publishes per-token pricing but no customer or revenue metrics
- Fine-tuning-as-a-service is a crowded, margin-thin category against hyperscalers and cheaper GPU clouds
Opportunities
- Interaction models could define a new product category (real-time, full-duplex, interruptible multimodal AI) - first-party FD-bench figures claim a latency lead over Gemini and GPT realtime stacks
- Open-weight model ecosystem (Llama, Qwen, DeepSeek, Kimi-class weights) keeps expanding, growing Tinker's addressable base
- Enterprise post-training/RL customization is early — Tinker's low-level Python API is differentiated vs. black-box fine-tuning endpoints
- A successful interaction-model launch later in 2026 could reopen the stalled mega-round at a defensible markup
Threats
- Frontier labs (OpenAI, Google DeepMind, Anthropic, Meta) can replicate both the fine-tuning product and low-latency multimodal interaction with far more compute and distribution
- Talent raids from Meta Superintelligence Labs and OpenAI directly target its core asset - both have already extracted co-founders
- Compute commitments (tens of billions reported for the NVIDIA supply arrangement) could outrun funding if the next raise stalls again - the Jan 2026 failure showed financing is not assured
- AI valuation regime change — the 2025-26 repricing of pre-revenue labs hits Thinking Machines hardest as the emblematic priced-for-perfection seed
Moats, dependencies & bottlenecks
Moats
John Schulman, Lilian Weng) Still elite, but the Oct 2025 - Jan 2026 departures (Tulloch, Zoph, Metz, Schoenholz) show it is contestable by Meta/OpenAI compensation
Google Cloud GB300 NVL72) Contracted advantage, but rivals hold larger absolute allocations
weak-to-moderate Real researcher goodwill and published pricing, but low switching costs (OpenAI-compatible API cuts both ways) and no disclosed commercial traction
sub-0.4s claimed latency) Research preview with first-party benchmarks only; incumbents can close latency gaps quickly
Dependencies
compute supplier + strategic investor 1+ GW Vera Rubin commitment reported at tens of billions; also an undisclosed-size investor - aligned but concentrated
cloud infrastructure Single-digit-billions partnership (Apr 2026) with a company that is also a direct competitor via DeepMind
Alibaba Qwen, DeepSeek, Moonshot Kimi) input for Tinker's product Tinker fine-tunes third-party open weights; license or release-policy changes upstream alter the catalog
No disclosed revenue with GW-scale commitments; Jan 2026 showed the next round is not guaranteed
leadership/talent The seed thesis is the team; further senior departures would be existential to the valuation
Advantages
- Largest seed war chest in venture history ($2B) at a stage where most rivals had a fraction
- Founding team that built ChatGPT-era post-training (RLHF) at OpenAI - credible in exactly the customization layer Tinker sells
- Two shipped artifacts in 15 months (Tinker GA — TML-Interaction-Small preview) plus respected public research on deterministic inference and LoRA post-training
- Strategic cap table (NVIDIA, AMD, Cisco, ServiceNow, Jane Street) that doubles as supplier, infrastructure, and enterprise channel relationships
Weaknesses
- No disclosed revenue or customer metrics for any product (pricing is published; traction is not)
- Valuation overhang from the collapsed ~$50B round; next raise anchors against a public failure
- Co-founder and senior-researcher attrition through H1 2026, including an acrimonious firing (Zoph)
- Strategic ambiguity — developer-infra business (Tinker) vs. frontier consumer-facing models (interaction models) compete for the same compute and talent
Bottlenecks
- Revenue conversion: published Tinker pricing but no disclosed paying-customer base as of 2026-07
- Financing: must re-open the growth round (or show revenue) before the $2B seed is consumed by GW-scale compute and elite payroll
- Product cadence — interaction models must go from announced research preview (May 2026) to broad release (late 2026) on schedule to sustain the narrative
- Talent retention against Meta Superintelligence Labs and OpenAI counter-recruiting
Top signals & trends
Top signals
Prospective backers declined the markup on product-record grounds; sharpest available external mark
The most supply-constrained player in AI chose to fund and supply them; FT put the chip supply at tens of billions
Second hyperscale compute anchor; also a hedge against NVIDIA concentration
Erosion of the core seed-round thesis, with public acrimony
Consistent shipping cadence and published pricing, even without disclosed revenue
Credible technical differentiation, but first-party benchmarks and preview-stage - monetization unproven
Trends
Grows Tinker's addressable base of customizable models
Directly caused the round collapse; capital now demands product proof
Exactly the interaction-models bet; also draws incumbent competition
Their allocation is secured, but cheaper compute also lowers rivals' barriers
Raises retention cost of their primary asset
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Vera Rubin accelerators, 1+ GW multi-year commitment (Mar 2026, deploying from early 2027); also strategic investor
Single-digit-billions AI Hypercomputer partnership incl. GB300 NVL72 / A4X Max access (Apr 2026); relationship since 2025
Seed investor and prospective alternative accelerator supplier
Tinker's early-adopter base (academic and safety-research groups fine-tuning open models, seeded via the 2026 interactivity grant program); named commercial customers not disclosed
Target segment for Tinker GA; no disclosed logos as of 2026-07
Private; the reference frontier lab, direct competitor in post-training and realtime voice; re-absorbed two TML co-founders plus a senior researcher in Jan 2026
Private; frontier lab with strong enterprise API business - the revenue traction TML lacks
Gemini Live attacks the same real-time interaction UX; Alphabet is also TML's cloud supplier
Open-weight Llama commoditizes the base-model layer; poached co-founder Andrew Tulloch
Private; competing frontier lab with massive owned compute (Colossus)
Private (France); open-weight lab with enterprise fine-tuning offerings
Private; open-model fine-tuning + inference cloud - Tinker's most direct infra competitor
Private; fast open-model inference/fine-tuning platform
Private (IPO candidate); enterprise model customization inside the data platform
Private; default hub + training stack for open-weight customization
Hyperscaler fine-tuning endpoints bundled with cloud spend
Enterprise distribution of fine-tuning + OpenAI models
Mainland-China open-weight labs - competitive context only, not a buy/own call: their releases simultaneously feed Tinker's catalog (Qwen3, Kimi K2 are hosted models) and compress the value of mid-scale proprietary models