Every time a new frontier model ships, the same question follows for the people who actually have to deploy it: how do you turn a general-purpose system into something that reliably does your specific job? That work — the fine-tuning, the alignment, the customization that happens after a model is pre-trained — is what the industry calls post-training. And it is quietly becoming one of the most contested layers in the AI stack.
This week, Bespoke Labs said it had raised $40 million to build infrastructure for exactly that layer. The company’s bet is straightforward, and increasingly popular among founders and investors: as base models commoditise, the value migrates to whoever can adapt them best. For anyone deploying AI — in India or anywhere else — it is worth understanding why.
The raise
Bespoke Labs announced $40 million in funding, structured across two stages. According to SiliconANGLE, roughly $31.75 million came via a Series A round led by Wing VC, with participation from Mayfield and The House Fund. That built on an earlier $8.25 million seed round; per the company’s own announcement, the seed was led by 8VC and included Google DeepMind chief scientist Jeff Dean among its backers. The cap table also reportedly includes angels who work at Anthropic, OpenAI and Meta — an unusual concentration of insiders from the very labs building the base models this startup sits on top of.
The company, founded in 2024 by Mahesh Sathiamoorthy (a former Google DeepMind researcher who is CEO) and Alex Dimakis (chief scientist, and a professor who has taught at UC Berkeley and UT Austin), builds tools for the post-training phase: creating the reinforcement-learning environments, data-curation pipelines and evaluation harnesses that turn a raw model into a dependable agent. In Sathiamoorthy’s framing, he left DeepMind with a goal to “democratize post-training” — the idea being that adaptation gets dramatically easier once you have good data and good environments to train in.
Bespoke is not starting from a blank page here. The team is already known for open projects including OpenThoughts (a reasoning dataset used by teams at Meta, Amazon and AI2), Terminal-Bench (an agentic-coding benchmark referenced by Anthropic, OpenAI and Google DeepMind), GEPA (an evolutionary optimiser for prompts and policies) and Curator, its synthetic-data-curation library. The new capital, the company says, goes toward scaling that infrastructure and expanding its research team.
Why post-training matters
To see why investors are writing cheques for the “layer after the model,” it helps to be precise about what post-training actually is. Pre-training is the expensive, capital-intensive part: feeding a model enormous quantities of text and code until it acquires general capabilities. Post-training is everything that comes after — supervised fine-tuning on curated examples, reinforcement learning from feedback, alignment, and increasingly, training inside simulated environments where an agent can practice a task and be rewarded for doing it well.
The strategic logic is about where differentiation lives. Base models are, more and more, a commodity. Several are competitive on general benchmarks, capable open-weight models are freely downloadable, and the gap between the best proprietary system and a good open one has narrowed for many everyday tasks. When the underlying engine is broadly available to everyone, owning the engine stops being a moat.
What is not commoditised is the adaptation. A model tuned on your proprietary data, aligned to your domain’s rules, and evaluated against your definition of “correct” is something a competitor cannot simply download. That is the case Bespoke is making: as the company puts it, the environment an agent learns in “is the only important component that is not going to be democratized.” You get real differentiation without the ruinous cost of training a base model from scratch — you rent the base and own the customization.
The competitive picture
This is not an empty field. The tooling layer around fine-tuning, data curation, evaluation and RL environments is crowded, with cloud providers, open-source frameworks and a growing crop of specialised startups all competing for the same budget. That is a signal of demand, not a warning against it — a layer this contested is a layer people believe is essential.
Two forces are inflating that demand. The first is the rise of capable open models: the more a team can legally take a strong open-weight model and shape it, the more they need reliable machinery to do the shaping. The second is the shift from chatbots to agents — systems that take multi-step actions in the world. Agents are far harder to make dependable than single-turn assistants, and that is precisely why training environments and rigorous evaluation have moved to the centre of the conversation. Bespoke’s own pitch leans on this: reliable agents, it argues, come less from a bigger base model and more from a better environment to learn in.
The through-line across all of it is reliability. Anyone shipping AI into production quickly discovers that raw capability is the easy part; the hard part is making the system behave consistently, safely and measurably. That is why evaluation — the unglamorous work of knowing whether your model actually got better — sits right alongside training in this stack.
The India read
For Indian teams, the post-training thesis is not an abstract American venture story — it maps almost directly onto where local value can be created. India’s advantage in AI was never going to come from spending billions to pre-train a rival to the largest frontier models. It comes from adaptation: taking capable base models and making them genuinely useful for Indian languages, scripts, regulatory contexts and industry-specific data.
That is fundamentally a post-training problem. Fine-tuning a model to handle Hindi, Tamil, Bengali or code-mixed text well; aligning it to the compliance requirements of Indian banking, healthcare or public services; curating high-quality domain data where none exists off the shelf — this is where a comparatively small, sharp team can add value that a generic global model cannot. And it is cost-effective in a way that base-model training never will be: adaptation is measured in the tens of thousands of dollars and weeks of work, not the hundreds of millions and years that pre-training demands.
The practical takeaway for founders and operators here is the same one the Bespoke raise underlines. If your AI strategy depends on owning a base model, you are competing on the most capital-intensive terrain in the industry. If it depends on owning the customization — the data, the fine-tuning, the evaluation of what “good” means for your users — you are competing where effort and domain knowledge still beat sheer scale. That is the layer after the model, and increasingly, it is where the real work is.
