Capital One Ventures Led a $150 Million Round Into Sentra to Solve the Problem AI Agents Create: They Touch Data You Didn't Know They Could Reach.

On November 6, 2026, Thinking Machines Lab did two things in the same day: it announced Inkling 2 Small, a new open-weight model whose weights will be released under the Apache 2.0 license within the week, and it became the subject of Business Insider reporting that the company is in advanced talks to close a $5 billion funding round at a $40 billion pre-money valuation, led by Accel.

The company declined to comment. The timing of the two announcements was likely not accidental.

Mira Murati, the lab's founder and former CTO of OpenAI, has spent 2026 building a company whose product thesis is almost the inverse of her former employer's: rather than building the most capable closed model and renting access to it, Thinking Machines Lab is building open-weight models designed from the ground up for enterprise customization, and monetizing the tooling — its Tinker platform — that makes customization practical.

The $5 billion raise, if it closes as reported, would value the company at roughly $45 billion post-money, making it one of only four privately held AI labs with a valuation above $30 billion. The jump from the $12 billion seed valuation in July 2025 is a 3.3x increase in roughly 16 months. For a company that has not disclosed revenue, it is a substantial bet on the thesis behind the product.

This piece explains what that thesis is, why it differs structurally from the OpenAI and Anthropic playbooks, and what the round's terms reveal about where investors believe enterprise AI value is headed.

The $40B Question: What Changed Since the Seed

When Thinking Machines Lab raised its $2 billion seed in July 2025, the company was pre-product and entirely thesis-driven. The bet was simple: Mira Murati's operational credibility from scaling OpenAI during its fastest growth period, combined with a founding team of senior AI engineers, was enough to anchor a $12 billion valuation before a single model was trained.

The $40 billion figure requires a more grounded explanation. In roughly 16 months, Thinking Machines has shipped Inkling (July 15, 2026), Inkling Small (July 31, 2026), and now Inkling 2 Small — a cadence that suggests functional infrastructure for large-model training. It has established the Tinker customization platform as a commercial product. And it has accumulated meaningful signal from enterprise deployments, though the company has not disclosed customer names or revenue.

What it has not done is claim the top spot on the EVALS leaderboard. When Inkling launched in July, Murati said explicitly that it was trained "for strong performance across the board," but that it is not the strongest model available today, closed or open. That framing — competitive but not dominant on benchmarks — is unusual in an industry where every frontier lab claims its latest release as the best model in the world.

The valuation multiple, then, reflects investor conviction that benchmark leadership is not the right metric for the enterprise AI market Thinking Machines is targeting. It is a bet that customization infrastructure is more valuable than raw performance. Whether that bet is correct will depend on how quickly the gap between frontier closed models and customizable open-weight models closes — and whether enterprises actually convert their stated preference for open-weight models into purchasing decisions.

Murati's Thesis: Customization Over Benchmarks

The core of Thinking Machines Lab's commercial argument is a critique of the current frontier AI market structure.

The existing market is built on access, not ownership. Enterprises pay OpenAI, Anthropic, or Google for API tokens and hosted inference. The model itself — its weights, its training data, its architectural choices — is a black box. When those enterprises want the model to perform better on their specific data, their specific vocabulary, their specific output format, they submit examples through few-shot prompting and hope the model generalizes. For some use cases, that is sufficient. For many enterprise use cases, it is not.

Regulated industries face an additional constraint: they cannot route sensitive data — patient records, cardholder data, financial analysis, legal work product — through third-party APIs. They need models they can run on their own infrastructure, in their own cloud, against their own data, with the assurance that none of that data is used for further model training. Open-weight models are the only architecture that provides this assurance, because there is no third-party involved after the weights are deployed.

Murati's bet is that these constraints are not niche requirements. They are the dominant requirements for the most valuable enterprise AI deployments: the ones in healthcare, financial services, legal, defense, and infrastructure that handle the most sensitive data and generate the highest willingness to pay. By building for those requirements from day one — open-weight models, enterprise-grade customization tooling, on-premises and private-cloud deployment options — Thinking Machines Lab is targeting the segment of the enterprise AI market that is structurally underserved by OpenAI and Anthropic's current closed-model approach.

The risk in the thesis is that the enterprise AI customization market develops more slowly than expected, or that the performance gap between customized open-weight models and closed frontier models remains large enough that regulated enterprises accept the compliance tradeoffs of using closed APIs through private deployment arrangements. Anthropic's revised S-1 filing suggests that the closed-model incumbents are not conceding the regulated enterprise market without a fight — they have invested heavily in private deployment options that allow some enterprises to run models in their own cloud environments without exposing training data.

Inkling's Technical Architecture — MoE, Not Monolith

The technical choice behind Inkling — mixture-of-experts rather than a dense transformer — is not arbitrary. It reflects a specific set of tradeoffs that matter for enterprise customization.

A dense transformer model, like GPT-4o in its standard configuration, routes every token through every parameter at inference time. The model is a single, monolithic system. Customizing it means fine-tuning the full parameter space, which is computationally expensive and risks catastrophic forgetting — where the model loses general capabilities as it learns domain-specific patterns.

Inkling's MoE architecture — 975 billion total parameters, 41 billion active at inference — works differently. At inference time, the model routes each token through a subset of specialized "experts," each of which has learned to handle specific types of content or tasks. At fine-tuning time, Tinker can selectively update specific experts relevant to the enterprise's domain while leaving the rest of the model intact. This approach reduces the risk of catastrophic forgetting, lowers the compute cost of customization, and allows multiple domain-specific variants to coexist in the same base architecture.

For enterprise customization at scale, the MoE architecture has a practical advantage over dense transformers: the customization cost scales with the number of relevant experts rather than the full model size. A healthcare system fine-tuning on electronic health record data doesn't need to update the full 975B parameter space. It updates the experts most activated by clinical documentation patterns, leaving the rest of the model unchanged. The result is a customized variant that is cheaper to produce and more stable across distribution shifts than a fully fine-tuned dense model of equivalent total parameter count.

Inkling 2 Small's Apache 2.0 release is a deliberate distribution play. Apache 2.0 is one of the most permissive open-source licenses — it allows commercial use, modification, and distribution without requiring derivative works to adopt the same license. By releasing under Apache 2.0, Thinking Machines Lab removes every legal barrier to enterprise adoption, including the usage restrictions in some competing open-weight licenses that have caused hesitation in regulated industries.

The Tinker Platform: Where the Business Actually Lives

Inkling is a distribution strategy. Tinker is the business.

The strategic logic mirrors Red Hat's model for the Linux ecosystem. Red Hat did not sell Linux — Linux was free and open source. Red Hat sold the enterprise subscription services that made Linux practical to run in production: certification, support, security patching, compatibility guarantees, and the professional services to implement it correctly. The business was valuable enough that IBM acquired Red Hat for $34 billion in 2019.

Thinking Machines Lab's version of this model is Tinker: the platform that makes Inkling practical for enterprise deployment at scale. Tinker handles the full lifecycle of enterprise model customization:

1. Data ingestion and preprocessing. Tinker connects to enterprise data systems — document repositories, CRM databases, EHR systems, internal knowledge bases — and processes them into training-ready formats. This step alone saves enterprise data science teams weeks of work per customization project, because domain-specific data formatting requirements are built in rather than custom-engineered.

2. Expert routing analysis. Before fine-tuning begins, Tinker analyzes which of Inkling's expert layers activate on the enterprise's data. This identifies the minimum set of experts to update, reducing compute cost and customization time.

3. Supervised fine-tuning with domain feedback. Tinker provides interfaces for domain experts — not just ML engineers — to review and annotate model outputs during the fine-tuning process. A radiologist can flag incorrect terminology in a healthcare deployment. A compliance officer can mark regulatory misstatements in a financial services deployment. This human feedback loop is integrated directly into the training pipeline.

4. Evaluation and benchmarking. After fine-tuning, Tinker generates domain-specific evaluation benchmarks to compare the customized variant's performance against the base model and against previous versions. This gives enterprises the evidence they need to validate that customization improved performance before deploying to production.

5. Deployment and version management. Tinker manages the deployment of customized model variants to on-prem, private cloud, or Thinking Machines Lab's managed cloud infrastructure. It handles versioning, rollbacks, and canary deployments, treating model updates with the same operational rigor as software deployments.

The revenue model for Tinker is not public, but the structure is likely a combination of platform subscription (access to Tinker infrastructure), consumption-based billing (for compute during fine-tuning and inference), and professional services for large enterprise customization projects.

How This Stacks Up Against OpenAI and Anthropic

OpenAI's commercial trajectory since the for-profit restructuring has moved steadily toward the enterprise, with ChatGPT Enterprise, OpenAI API, and the Operator tier all targeting large organizational buyers. The commercial model remains closed: enterprises pay for access to OpenAI's models, and their usage data may inform model improvements under various terms of service conditions.

Anthropic's revised S-1 filing describes a similar structure — closed models accessed via API, with private deployment options available for regulated industries under specific agreements. Anthropic has invested in infrastructure that allows some enterprises to run model inference in their own cloud environments, but the model weights remain Anthropic's proprietary assets.

Thinking Machines Lab is betting that this closed-model structure is a liability, not an asset, as enterprise AI matures. The argument is structural:

Data sovereignty. Enterprises with sensitive data requirements — covering an estimated 60-70% of large-enterprise AI spending by value — cannot accept API-based access models that require routing that data through a third-party's infrastructure. Open-weight models with on-prem deployment are the only architecture that eliminates this constraint entirely.

Customization depth. For high-value enterprise tasks — clinical documentation summarization, legal contract analysis, financial report generation — a model fine-tuned on the organization's specific data, processes, and terminology outperforms a general-purpose frontier model by a margin that justifies the customization cost. The more specialized the task, the larger the performance gap.

Vendor dependency risk. Enterprises building critical workflows on closed API models are dependent on the API provider's pricing, availability, and policy decisions. The fal.ai $1.65B inference round reflects the same concern at the infrastructure layer — enterprises and developers want model access that isn't entirely controlled by the labs that also compete with them in applications.

Long-term cost structure. At scale, the per-token economics of running a fine-tuned open-weight model on owned or reserved infrastructure outperform cloud API pricing by a significant margin. The break-even point depends on volume, but for large-enterprise deployments above a few million API calls per month, the economic case for in-house model deployment is increasingly compelling.

The Open-Weight Gambit

Thinking Machines Lab is not the first to bet on open-weight models as an enterprise AI strategy. Meta's Llama ecosystem has demonstrated that open-weight release generates distribution and ecosystem development at a scale that proprietary models cannot replicate organically. DeepSeek's R1 release demonstrated that open-weight models can be competitive with frontier closed models on specific benchmarks.

What Thinking Machines Lab adds to this playbook is the customization infrastructure layer — the enterprise tooling that converts a capable open-weight base model into a production-ready enterprise AI system. That layer is where the value creation happens for regulated enterprises, and it is structurally difficult for Meta (a consumer company) or DeepSeek (a research lab without an enterprise go-to-market) to deliver.

The $40 billion valuation implies that investors believe Thinking Machines Lab can capture a defensible position in this layer before the larger incumbents — Microsoft, Google, Salesforce — build equivalent customization tooling on top of their existing enterprise relationships. The race is real: Microsoft already offers Azure AI Fine-Tuning, Google offers Vertex AI Model Customization, and both have enterprise relationships that Thinking Machines Lab will need years to match.

The Apache 2.0 release of Inkling 2 Small's weights this week is an acceleration of that race. By putting the weights in the hands of enterprise data science teams immediately — before Microsoft or Google offers an equivalent open-weight MoE model with comparable customization tooling — Thinking Machines Lab creates distribution and community lock-in that is difficult to dislodge even after competing platforms arrive.

The sovereign AI and national LLM race adds an international dimension to this strategy: governments and enterprises outside the United States that require data sovereignty guarantees that neither OpenAI nor Anthropic can fully provide are a natural second-wave market for Thinking Machines Lab's open-weight, on-prem-deployable approach.

What Enterprise Buyers Should Watch

For enterprise technology buyers evaluating AI strategy in the context of Thinking Machines Lab's raise, several near-term indicators will determine whether the lab's thesis plays out:

The Inkling 2 Small Apache 2.0 release quality. When the weights drop this week, the enterprise data science community will immediately benchmark them against Llama 3.3 70B, DeepSeek R1, and other open-weight alternatives. If Inkling 2 Small demonstrates meaningful performance advantages on domain-specific benchmarks — not general leaderboards, but realistic enterprise tasks — it will accelerate Tinker adoption significantly.

Tinker's performance on regulated-industry use cases. The first production case studies from healthcare, financial services, or legal deployments will be the proof-point that justifies Tinker's pricing versus in-house fine-tuning using open-source tooling. If Tinker demonstrably outperforms DIY customization in production reliability, evaluation rigor, and deployment efficiency, the business case for the platform subscription becomes straightforward.

Competitive response from Microsoft, Google, and Salesforce. Each of these incumbents has the enterprise relationships that Thinking Machines Lab lacks. If they launch credible open-weight customization platforms before Thinking Machines Lab has established deep enterprise accounts, the lab's competitive window narrows. The Apache 2.0 release is a move to establish distribution before that competition materializes.

Team stability. The reported departure of roughly one-third of the founding team is a yellow flag. The lab has not disclosed who departed or for what reasons, but founding team continuity is a material factor in whether the technical roadmap is executable. Murati's credibility carries the company's reputation externally; the internal technical team executes the product.

The Broader Signal

Thinking Machines Lab's $5B raise, alongside Anthropic's $33B S-1 and revenue figures and the ongoing consolidation of enterprise AI spending reported by Menlo Ventures around a small number of dominant models, is clarifying a bifurcation in the frontier AI market that will determine which companies capture the largest share of enterprise AI spending over the next three years.

The first path is the closed-model incumbent path: build the strongest possible general-purpose model, distribute it through API and consumer products, and extract margin from both. OpenAI and Anthropic are on this path, and their current revenue run rates and valuation premiums reflect investor confidence in it.

The second path is the open-weight enterprise customization path: build strong-enough open-weight models, release them permissively to generate distribution, and monetize the customization and deployment infrastructure layer. Thinking Machines Lab, Meta's Llama ecosystem, and a growing constellation of enterprise AI infrastructure vendors are on this path.

The critical unknown is where the enterprise AI customization market sits in its maturity curve. If enterprises are 12-18 months from large-scale custom model deployment in production, Thinking Machines Lab's timing is nearly perfect — it has established the platform before most enterprise buyers have locked in customization infrastructure choices. If the market develops faster than that, the Microsoft and Google incumbents' head start in enterprise relationships may be enough to block the lab's expansion.

Takeaway: Thinking Machines Lab's $5B raise at a $40B valuation is not primarily a bet on Inkling's current benchmark performance. It is a bet on a market structure thesis: that enterprise AI value will migrate from which model scores highest on EVALS to which model fits most precisely into an organization's specific regulatory constraints, data architecture, and domain requirements. If that thesis is correct, the lab's open-weight, customization-first architecture is structurally better suited to capturing the highest-value enterprise AI deployments than the closed-model incumbents. The Apache 2.0 release of Inkling 2 Small this week is the opening move in a race to establish distribution before that market structure question is settled.