Thinking Machines Lab Public-Source Startup Diligence Report
The investable thesis is that a high-credibility AI founding team can convert Tinker, open-model customization workflows, and a large NVIDIA compute partnership into a differentiated platform for enterprise, research, and developer customization. The central diligence question is whether this can become a durable, margin-positive business before the $12B valuation and compute obligations outrun product-market-fit evidence.
Thinking Machines Lab Public-Source Startup Diligence Report
Thinking Machines Lab passes the public eligibility screen for a startup diligence target: it is publicly described as privately held, has no SEC public-company ticker match, appears active through product, hiring, and partnership announcements, and has public independent reports of a $2B seed round at a $12B valuation. The diligence view is constructive but high-risk: elite team and funding signals are strong, while revenue quality, customer economics, cap table, compute commitments, and legal/IP position are not publicly verifiable.
Public sources do not disclose audited or management financials, ARR, usage revenue, gross margin, burn, runway, backlog, or AR aging.
Diligence request: Request inception-to-date financials, board plan, revenue schedules, billings, cash burn, runway, AR aging, and backlog.
Tinker is young and was launched as a private beta; no public paid conversion, uptime, retention, or gross-margin metrics are available.
Diligence request: Request product usage cohorts, uptime, job success rates, support tickets, paid conversion, retention, and product P&L.
Cloud AI vendors, frontier labs, and open-source stacks can compete on model quality, price, distribution, and enterprise trust.
Diligence request: Request win/loss data, benchmark comparisons, pricing analysis, switching-cost evidence, and roadmap differentiation.
A $12B valuation shortly after launch requires rapid commercial proof, product reliability, and defensible differentiation.
Diligence request: Review financing model, milestone plan, investor materials, valuation comps, and downside scenarios.
The announced one-gigawatt Vera Rubin deployment could create material supplier, capex, delivery, utilization, and contractual dependence.
Diligence request: Review NVIDIA agreements, deployment milestones, minimum commitments, financing terms, alternative suppliers, and utilization model.
Public sources name pilots but not top customers, revenue concentration, churn, contracts, or satisfaction.
Diligence request: Run reference calls and request top-customer schedules, contracts, usage, NPS, renewal, churn, and expansion data.
Company credibility heavily depends on a small set of prominent founders and senior AI researchers; compensation and turnover data are private.
Diligence request: Review employment agreements, equity vesting, retention plans, attrition, references, and succession planning.
Frontier AI and real-time multimodal systems can trigger privacy, safety, misuse, export-control, and emerging AI regulatory obligations.
Diligence request: Review AI safety governance, red-team results, regulator correspondence, export-control analysis, incident logs, and release criteria.
Public financial visibility is thin. The company has credible public funding and valuation evidence, but no audited statements, revenue, backlog, gross-margin, burn, or runway data. Financial diligence should focus on whether product usage and commercial traction can justify the valuation and any NVIDIA-linked infrastructure obligations.
not publicly verifiable confidence: high
No public income statements, balance sheets, cash-flow statements, management reports, sales breakdowns, backlog, or AR aging were found.
| information type | public status | diligence implication |
|---|---|---|
| Income statements, balance sheets, cash flows, footnotes | Not found | Cannot assess revenue, burn, runway, gross margin, or working capital from public sources. |
| Planned versus actual results and management reports | Not found | Forecast credibility and execution variance are not publicly testable. |
| Sales/gross profit by product, channel, geography, backlog, AR aging | Not found | Customer quality, revenue concentration, and collections risk remain open diligence items. |
Public sources confirm funding and product activity but not standard operating financials.
partially verified confidence: medium
Public sources identify growth drivers but not board-approved projections or sensitivity cases.
| financial area | public signal | risk | request |
|---|---|---|---|
| Revenue quality | Tinker launched in private beta with usage-based pricing planned. | No proof of paid recurring revenue, gross margin, or retention. | Product P&L, usage cohorts, pricing contracts, and customer invoices. |
| Compute commitments | NVIDIA partnership targets at least one gigawatt of Vera Rubin systems. | Potential capex, lease, services, or take-or-pay obligations could dominate burn. | NVIDIA agreement, deployment plan, utilization model, and accounting treatment. |
| Capital structure | $2B seed at $12B valuation and later NVIDIA investment. | Unknown dilution, preference stack, governance rights, and option-pool needs. | Pro forma cap table, stock plan, financing documents, and investor rights agreement. |
Financial diligence is the highest-priority private-data request.
Valuation is reported by media and list sources, not verified from financing documents.
partially verified confidence: high
Round size, valuation, lead investor, and several participants are public, but share count, option pool, debt, warrants, liquidation preferences, and investor rights are not.
| date | event | public terms | unresolved terms |
|---|---|---|---|
| 2025-02-18 | Company launch coverage | Company founded by Mira Murati and launched with public leadership team; no financing terms in launch article. | Formation capitalization, founder ownership, and option pool. |
| 2025-07-15 | Seed financing | $2B seed round at reported $12B valuation led by Andreessen Horowitz, with NVIDIA, AMD, Accel, ServiceNow, Cisco, and Jane Street among reported participants. | Security type, liquidation preference, board rights, ownership, debt, and side letters. |
| 2026-03-10 | NVIDIA strategic investment and partnership | Significant investment plus multiyear partnership targeting at least one gigawatt of Vera Rubin systems. | Investment amount, security, compute purchase commitments, exclusivity, and payment schedule. |
Public values should be reconciled to signed financing documents.
Dates are public-source event dates, not necessarily legal closing dates.
not publicly verifiable confidence: high
Tax positions, accounting policies, revenue recognition, and detailed financing history were not publicly available beyond reported financing events.
The public product story centers on Tinker, a developer/researcher API and SDK for fine-tuning open-weight language models, plus broader research previews toward interactive, customizable AI systems. Product evidence is technically detailed, but maturity, uptime, retention, paid conversion, and enterprise security posture need private validation.
partially verified confidence: high
Tinker is the only clearly launched product found publicly; the broader platform and interaction-model work remain research or roadmap signals.
| product or artifact | public status | target users | key evidence |
|---|---|---|---|
| Tinker | Launched October 2025 in private beta | Researchers and developers fine-tuning language models | API primitives, docs, pricing/data statements, and cookbook examples are public. |
| Interaction models | Research preview | Future human-AI collaboration experiences across audio, video, and text | Official post describes native interaction models, micro-turn design, real-time multimodal interaction, and safety work. |
| Frontier/customizable AI systems | Company mission and partnership direction | Enterprises, research institutions, scientific community, and advanced AI users | Official website and NVIDIA partnership describe customizable AI at scale and broader access. |
Tinker is the only publicly launched product identified.
| dimension | public claim | diligence test |
|---|---|---|
| Training workflow | Tinker exposes forward/backward, optimizer-step, sampling, and state-management primitives through an SDK. | Benchmark reliability, throughput, queueing, job failure rates, and developer experience. |
| Pricing and supported models | Public Tinker materials include a pricing table and support for open model families such as Qwen, Llama, DeepSeek, GPT-OSS, Kimi, and Nemotron. | Review actual customer pricing, discounts, COGS, margin by model, and model-license compliance. |
| Data use | Tinker data is used solely to fine-tune user models and not to train Thinking Machines Lab's own models. | Review DPA, retention, access controls, subprocessor list, audit logs, and deletion workflows. |
Product commitments should be compared against signed customer terms.
Architecture is inferred from public docs and README, not internal design docs.
Named public customer evidence is limited to Tinker pilots and strategic relationships. No top-15 customer list, revenue by customer, churn, backlog, or supplier spend schedule is public. NVIDIA is both a strategic investor/partner and likely critical supplier, creating concentration and dependency diligence priorities.
partially verified confidence: medium
Public materials name academic and research pilot users, not top customers by revenue or application.
| organization | relationship type | public evidence | missing commercial detail |
|---|---|---|---|
| Princeton, Stanford, Berkeley, Redwood Research | Tinker pilot users | Official Tinker announcement listed them as users. | Payment status, contract value, usage, renewal, and satisfaction. |
| NVIDIA | Strategic investor, partner, and compute supplier | Official company and NVIDIA announcements describe multiyear partnership, at least one gigawatt of Vera Rubin systems, and significant investment. | Investment amount, supplier spend, purchase commitments, exclusivity, and delivery schedule. |
| Andreessen Horowitz, AMD, Accel, ServiceNow, Cisco, Jane Street | Reported investors | Independent funding coverage named a16z as lead and reported strategic/financial investor participation. | Ownership, governance, commercial rights, and strategic obligations. |
No customer revenue by organization is public.
| requested item | public status | risk if unresolved | requested evidence |
|---|---|---|---|
| Top 15 customers and applications | Not available; only selected pilots named. | Unknown concentration, buyer profile, and mission-criticality. | Customer list, contracts, usage logs, and application taxonomy. |
| Revenue by customer and greater-than-5-percent accounts | Not available. | Cannot assess customer concentration or revenue predictability. | Revenue and billings schedule by customer for inception to date and current forecast. |
| Severed relationships | No public evidence found; inconclusive. | Failed pilots or partner disputes could be hidden. | Lost customer, churn, dispute, and supplier-change schedule. |
Customer diligence should prioritize reference calls and contract review.
Counts reflect organizations named in public sources reviewed.
partially verified confidence: high
The strongest public strategic relationship is NVIDIA; investor and pilot relationships are also visible, but marketing and revenue-contribution terms are not.
not publicly verifiable confidence: high
No customer revenue schedule or greater-than-5-percent customer analysis is publicly available.
inconclusive confidence: medium
No public evidence of severed customer, partner, or supplier relationships was found in this pass.
partially verified confidence: high
NVIDIA is the only public top-supplier-level dependency; other cloud, data-center, networking, data, and model suppliers are not publicly disclosed.
| supplier or dependency | public evidence | diligence risk |
|---|---|---|
| NVIDIA compute systems | Multiyear partnership to deploy at least one gigawatt of next-generation Vera Rubin systems. | Concentration, delivery timing, cost, contract terms, and utilization risk. |
| Open model families supported by Tinker | Tinker page references support for model families including Qwen, Llama, DeepSeek, GPT-OSS, Kimi, and Nemotron. | License compliance, model availability, performance variability, and dependency on third-party roadmaps. |
| Data-center, cloud, networking, and security vendors | Not publicly disclosed. | Hidden commitments, outages, security posture, and vendor lock-in. |
NVIDIA should be treated as a critical supplier and strategic counterparty.
Thinking Machines Lab competes in frontier AI systems, developer model-customization tooling, and enterprise AI infrastructure. Public technical and team signals are strong, but competitive differentiation is not yet proven through public market share, scaled customers, or pricing power.
partially verified confidence: medium
The company appears positioned at the intersection of AI lab, model-customization platform, and infrastructure-scale AI systems; competitive moats are plausible but unproven.
| segment | likely competitors | thinking machines position | diligence test |
|---|---|---|---|
| Fine-tuning and post-training APIs | Cloud AI platforms, frontier-lab APIs, and open-source training stacks. | Tinker provides SDK-level training primitives and cookbook recipes for advanced users. | Win/loss, pricing, speed, quality, and retention versus alternatives. |
| Frontier customizable AI systems | Frontier AI labs and cloud model platforms. | Company mission and NVIDIA partnership point to frontier training and customizable AI at scale. | Model benchmarks, safety tests, data advantage, and enterprise procurement readiness. |
| Real-time interaction models | Real-time voice/video model providers and multimodal AI labs. | Research preview claims native interaction through continuous multimodal micro-turns. | Latency, safety, benchmark validity, and productization timeline. |
Competitor names are archetypal in this public report because no company win/loss dataset was available.
| basis | supportive public evidence | risk |
|---|---|---|
| Technical depth | Published research on LoRA, deterministic inference, distillation, and interaction models. | Research may be imitated or not productized into durable paid advantage. |
| Distribution and community | Tinker docs and open-source cookbook reduce adoption friction. | Community usage may not convert to enterprise revenue. |
| Compute scale | NVIDIA partnership targets at least one gigawatt of Vera Rubin systems. | Competitors may have equal or larger compute plus stronger distribution; underutilization could hurt margins. |
Pricing power is not established publicly.
Competitor placement is analyst inference, not sourced market share.
Public GTM evidence suggests an expert-led developer/researcher motion: private beta, docs, open-source cookbook, named pilots, and active hiring in GTM/product roles. The company has not publicly disclosed sales productivity, pipeline, quotas, or budget adequacy.
partially verified confidence: medium
The public strategy emphasizes developer/researcher access, open-source examples, private beta, and technical content marketing rather than broad enterprise sales proof.
| channel | evidence | positioning | risk |
|---|---|---|---|
| Private beta and product sign-up | Tinker launched in private beta for researchers and developers. | Advanced post-training API for expert users. | Beta usage may not convert to paid production workloads. |
| Open-source cookbook and docs | GitHub README describes tinker and tinker-cookbook plus tutorials and recipes. | Community and developer education path. | Open-source adoption does not prove proprietary monetization. |
| Strategic partnership visibility | NVIDIA partnership announcement highlights enterprises, research institutions, and scientific community. | Credible frontier compute and platform scale. | Strategic dependence or exclusivity may constrain broader distribution. |
No sales productivity or CAC data is public.
Only named pilots have a public count.
partially verified confidence: medium
Major customer status and pipeline are not public; named pilots are the only visible demand signal.
| area | public signal | missing private data |
|---|---|---|
| Pipeline | Named pilots and beta access; no pipeline amount. | Pipeline by stage, ACV, probability, sales cycle, and conversion history. |
| Sales force | Greenhouse shows GTM/product roles, but no sales org metrics. | Quota capacity, compensation, ramp, win rate, and productivity by rep. |
| Budget | $2B seed provides funding capacity; NVIDIA compute may require substantial budget. | Board budget, hiring plan, compute capex/opex, and marketing allocation. |
Sales diligence cannot be completed from public sources alone.
partially verified confidence: medium
Public new-business channels appear to be beta sign-ups, technical documentation, GitHub/community distribution, research publications, and strategic partner visibility.
not publicly verifiable confidence: high
No public sales compensation, quota, productivity, or sales-cycle model was found.
inconclusive confidence: medium
The July 2025 seed round suggests substantial funding, but budget allocation and plan feasibility are not public.
R&D is the strongest public evidence category. The company has published detailed technical work, shipped Tinker with a cookbook, and announced an NVIDIA partnership to support frontier training. The main diligence gap is translating technical output into proprietary advantage, product economics, and safe deployment.
partially verified confidence: high
Public evidence shows a research-heavy organization with prominent AI leaders, frequent technical posts, open-source implementation examples, and active technical hiring.
| publication or asset | topic | diligence relevance |
|---|---|---|
| LoRA Without Regret | Efficient fine-tuning and low-rank adaptation performance conditions. | Supports Tinker fine-tuning thesis and cost-efficiency narrative. |
| Defeating Nondeterminism in LLM Inference | Batch-invariant and reproducible inference behavior. | Relevant to reliability, debugging, reproducible training/evaluation, and enterprise trust. |
| On-Policy Distillation | Dense teacher feedback for student model training, implemented with Tinker. | Demonstrates Tinker use for advanced post-training and potential efficiency advantages. |
| Interaction Models | Native real-time multimodal human-AI collaboration. | Signals future product direction beyond Tinker. |
Technical review should examine reproducibility and unpublished workpapers.
Counts are outputs reviewed in this diligence pass, not total company research output.
partially verified confidence: medium
Public pipeline signals include Tinker expansion, interaction models, and NVIDIA-backed frontier training; development costs and milestone confidence are not public.
| pipeline item | public status | critical dependency | key risk |
|---|---|---|---|
| Tinker expansion | Private beta with docs, pricing page, and cookbook. | Reliable distributed training service, supported models, and customer support. | Product-market fit, uptime, margins, and enterprise security. |
| Interaction models | Research preview. | Real-time multimodal training, inference optimization, safety, and UX. | Productization timeline, safety, latency, and user demand. |
| Frontier model training/platform | NVIDIA partnership announced; deployment targeted for early next year. | Vera Rubin system delivery, data-center capacity, financing, and utilization. | Supplier concentration and infrastructure obligations. |
Pipeline status should be reconciled to management roadmap.
Public leadership credibility is high, and hiring signals indicate rapid organizational buildout. However, org structure, compensation, stock plans, turnover, employee relations, and retention risk are not publicly verifiable.
partially verified confidence: medium
Public materials identify senior leaders and functional hiring areas, but no complete organization chart is public.
| person | role | public basis | diligence need |
|---|---|---|---|
| Mira Murati | Cofounder and CEO | Official/company and launch coverage identify her as founder/CEO. | Employment terms, vesting, references, prior-employer obligations, and key-person planning. |
| John Schulman | Cofounder/chief scientist | Launch coverage identifies him as chief scientist. | Research roadmap ownership, retention terms, and invention assignment. |
| Barret Zoph | Cofounder/CTO | Launch coverage identifies him as CTO. | Infrastructure/product execution scope, retention terms, and invention assignment. |
The report does not include non-essential employee-level personal data.
Reporting lines other than public roles are inferred and require management validation.
partially verified confidence: medium
Public sources show launch headcount, LinkedIn profile counts, and active roles, but not verified payroll headcount or hiring plan.
| signal | value or observation | interpretation |
|---|---|---|
| Launch headcount | 29 employees reported at launch. | Confirms early team buildout but not current payroll size. |
| LinkedIn profile/company-size data | LinkedIn showed 177 associated profiles and an 11-50 company-size field. | Useful activity signal but inconsistent/stale for diligence-grade headcount. |
| Greenhouse open roles | Roles across legal, HR, GTM, product/design, engineering, research, and infrastructure. | Indicates rapid functional buildout centered on San Francisco hiring. |
Payroll records are required for verified headcount.
Payroll headcount is needed for diligence-grade analysis.
partially verified confidence: high
Public launch coverage names senior leaders and prior AI affiliations, but complete biographies, board roles, and employment terms need diligence.
partially verified confidence: medium
Public job postings include salary and benefits signals for at least some roles, but executive compensation and employment agreements are not public.
| people area | public signal | hidden risk | diligence request |
|---|---|---|---|
| Compensation | Designer posting showed $350k-$475k salary range plus benefits and visa sponsorship. | High compensation may accelerate burn and set expensive internal benchmarks. | Payroll, offer letters, comp bands, bonuses, contractor spend, and benefits plan. |
| Incentive stock | No public stock plan found; large valuation and hiring imply importance. | High strike prices, option-pool depletion, and retention gaps. | Option plan, grants, strike prices, vesting, refresh policy, and fully diluted pool. |
| Employee relations and turnover | No public litigation found; no turnover data public. | Prior-employer disputes, rapid-hiring friction, or attrition may not be visible. | HR complaint log, attrition, retention, exit interview, and threatened-claim schedule. |
Personnel data requires management and counsel support.
not publicly verifiable confidence: high
No public stock-option plan, grant schedule, strike prices, or vesting provisions were found.
inconclusive confidence: medium
No public employee-relations disputes were found, but no internal HR records were available.
not publicly verifiable confidence: high
Personnel turnover is not publicly verifiable; public profile counts are too noisy for attrition analysis.
Public searches did not identify litigation or public-company status, and the cookbook is Apache-2.0 licensed. However, legal diligence remains high-priority because frontier AI implicates IP ownership, training-data rights, privacy/security, regulatory obligations, supplier contracts, and insurance exposures that are not public.
partially verified confidence: medium
CourtListener API search returned no results for "Thinking Machines Lab" in this pass.
| search area | method | result | limitation |
|---|---|---|---|
| Litigation | CourtListener API exact-name search. | Count 0 for "Thinking Machines Lab". | Does not cover every state court, arbitration, sealed matter, affiliated entity, or threatened claim. |
| Public-company status | SEC company_tickers JSON name search. | No public-company ticker match found. | Does not rule out private securities filings or affiliated entities. |
| Regulatory agency problems | Public-source review of available news/company sources. | No agency problem identified in this pass. | Not an exhaustive federal, state, international, or paid-database search. |
Counsel should perform docket, corporate, IP, employment, and regulatory searches.
Heatmap is an analyst prioritization, not a legal opinion.
partially verified confidence: medium
No public CourtListener results were found indicating company-initiated lawsuits.
inconclusive confidence: medium
No public workplace-safety or environmental liabilities were found, but compute infrastructure and frontier AI safety create operational and regulatory exposure.
| exposure or contract | public evidence | potential exposure | requested document |
|---|---|---|---|
| NVIDIA strategic partnership | At least one gigawatt Vera Rubin deployment target plus significant investment. | Minimum commitments, delivery failures, exclusivity, liability, and underutilized capacity. | Master agreement, capacity schedule, purchase/lease terms, side letters, and investment documents. |
| Customer data contracts | Public Tinker data commitment. | Privacy, confidentiality, data misuse, deletion, subprocessor, and audit rights. | Customer terms, DPA, privacy policy, retention schedule, and security audit. |
| AI safety and interaction-model deployment | Interaction models post discusses safety for real-time systems. | Misuse, harmful outputs, regulatory scrutiny, and incident response. | Safety policy, red-team logs, incident response plan, and model release criteria. |
Insurance coverage should be mapped to each exposure.
partially verified confidence: medium
The public cookbook is Apache-2.0 licensed, but patent, trademark, proprietary-model, training-data, and invention-assignment positions were not fully verified.
| area | public evidence | diligence need |
|---|---|---|
| Cookbook license | GitHub repository metadata and license indicate Apache-2.0. | Confirm ownership, contributor agreements, and third-party dependencies. |
| Research outputs | Official posts publish technical methods across LoRA, inference determinism, distillation, and interaction models. | Patent/trademark strategy, invention assignments, and proprietary components not disclosed. |
| Customer data and model rights | Tinker page states user data is solely for fine-tuning user models and not for company model training. | Contract language, data retention, subprocessor, audit, and deletion controls. |
Official patent and trademark registry checks remain a gap.
| legal area | public status | priority request |
|---|---|---|
| Corporate and securities | Private-company status supported by LinkedIn and SEC ticker search; financing terms private. | Charter, bylaws, stockholder agreements, financing documents, board minutes, and cap table. |
| IP and open source | Cookbook Apache-2.0 visible; patent/trademark portfolio not verified. | IP register, OSS scan, invention assignments, trademark searches, and patent filings. |
| Privacy, safety, and regulatory | Public data commitment and safety research statements; no regulatory problems found. | DPA, privacy/security controls, AI safety governance, export-control analysis, and regulator correspondence. |
Counsel should own this request list.
not publicly verifiable confidence: high
Insurance coverage was not publicly available.
partially verified confidence: high
Public material contracts include clear evidence of an NVIDIA strategic partnership and public Tinker data commitments, but actual agreements are not public.
inconclusive confidence: medium
No public regulatory agency problems were found in this pass, and no SEC public-company record matched; nevertheless frontier AI and customer-data handling merit regulatory diligence.
| ID | Claim | Status | Sources |
|---|---|---|---|
| EC-001 | Thinking Machines Lab publicly describes itself as an AI research and product company building customizable, generally capable AI systems. | verified high | SRC-001 |
| EC-002 | Public eligibility screen supports treating Thinking Machines Lab as a private unicorn, not a public company. | partially verified high | SRC-004SRC-006SRC-014SRC-023 |
| EC-003 | Thinking Machines Lab reportedly raised a $2B seed round at a $12B valuation in July 2025, led by Andreessen Horowitz with notable strategic and financial investors. | verified high | SRC-004SRC-005SRC-006 |
| EC-004 | Public launch coverage identifies Mira Murati as CEO, John Schulman as chief scientist, Barret Zoph as CTO, and 29 employees at launch. | verified high | SRC-003 |
| EC-005 | Tinker launched as a flexible API for fine-tuning language models, with SDK primitives for training and sampling. | verified high | SRC-007SRC-008SRC-009SRC-010 |
| EC-006 | Public Tinker materials include supported model families, pricing information, and a data-use commitment. | partially verified medium | SRC-008 |
| EC-007 | Thinking Machines Lab publicly named Princeton, Stanford, Berkeley, and Redwood Research as Tinker pilot users. | partially verified medium | SRC-007 |
| EC-008 | The tinker-cookbook repository provides open-source recipes, tutorials, and examples for model customization using Tinker. | verified high | SRC-010 |
| EC-009 | Public job postings show active hiring across legal, HR, GTM, product/design, engineering, research, and infrastructure roles. | verified high | SRC-012 |
| EC-010 | LinkedIn identifies the company as privately held, San Francisco based, with inconsistent public headcount/profile signals. | partially verified medium | SRC-014 |
| EC-011 | Thinking Machines Lab and NVIDIA announced a multiyear partnership targeting at least one gigawatt of NVIDIA Vera Rubin systems and a significant NVIDIA investment. | verified high | SRC-015SRC-016SRC-017 |
| EC-012 | The LoRA research post reports experiments suggesting LoRA can match full fine-tuning under certain post-training conditions. | verified medium | SRC-018 |
| EC-013 | The nondeterminism research post addresses reproducible LLM inference and batch-invariant behavior. | verified medium | SRC-019 |
| EC-014 | The on-policy distillation post describes using Tinker to combine on-policy relevance with dense teacher feedback for post-training. | verified medium | SRC-020SRC-010 |
| EC-015 | Thinking Machines Lab published a research preview of interaction models for native real-time multimodal human-AI collaboration. | verified medium | SRC-021 |
| EC-016 | No public operating financial statements, revenue, backlog, margin, burn, or AR data were found in reviewed sources. | not publicly verifiable high | SRC-001SRC-004SRC-005SRC-007SRC-008 |
| EC-017 | CourtListener exact-name search returned no cases for "Thinking Machines Lab" in this pass. | partially verified medium | SRC-022 |
| EC-018 | SEC company_tickers search found no public-company match for Thinking Machines Lab. | partially verified high | SRC-023 |
| EC-019 | The tinker-cookbook repository is licensed under Apache-2.0. | verified high | SRC-011 |
| EC-020 | Thinking Machines Lab competes in fine-tuning APIs, frontier AI systems, real-time multimodal interaction, and AI infrastructure/platform segments. | partially verified medium | SRC-001SRC-007SRC-015SRC-021 |
| EC-021 | Tinker product maturity is early; public launch materials described private beta and free-to-start usage with pricing later. | partially verified high | SRC-007 |
| EC-022 | Public GTM evidence points to technical content, docs, open-source cookbook distribution, private beta access, and hiring, not a fully disclosed sales motion. | partially verified medium | SRC-007SRC-010SRC-012 |
| EC-023 | The NVIDIA partnership creates a material compute supplier and infrastructure dependency. | verified high | SRC-015SRC-016SRC-017 |
| EC-024 | No public regulatory agency problem was found in this pass, but frontier AI, customer data, and interaction models create meaningful regulatory and safety exposure. | inconclusive medium | SRC-001SRC-008SRC-021SRC-022SRC-023 |
| EC-025 | Public sources do not disclose customer revenue, top customers, churn, backlog, or concentration. | not publicly verifiable high | SRC-007SRC-008 |
| EC-026 | Public sources do not disclose detailed capital structure, shares outstanding, stockholder list, options, warrants, notes, or debt. | not publicly verifiable high | SRC-004SRC-005SRC-015SRC-016 |
| EC-027 | Public Tinker data commitment states user data is only used to fine-tune user models, not train company models. | partially verified medium | SRC-008 |
| EC-028 | A public Designer job posting listed a $350k-$475k salary range, visa sponsorship, and benefits. | verified high | SRC-013 |
| EC-029 | Public sources show the company is active after launch through product, research, hiring, and partnership announcements. | verified high | SRC-001SRC-007SRC-010SRC-012SRC-015SRC-021 |
| EC-030 | Several requested diligence areas were not publicly verifiable in this pass. | not publicly verifiable high | SRC-022SRC-023 |
This report is a public-evidence diligence snapshot, not investment advice. Important financial, legal, technical, and contractual facts remain non-public and should be verified directly with management and primary documents before any investment decision.