Why we stopped chasing big models - and started building a platform

In August 2024, at Adya, we did what almost every ambitious AI company in the world was trying to do - we built our own large language model. At that moment, it felt like the right move.

The world was still in the GPT-4 era. Open-source models were improving rapidly. Fine-tuning promised real gains. And for our specific use case in retail and commerce, our early results were encouraging. By September 2024, our fine-tuned model was showing nearly 21% better task efficiency than the best models available at the time, while being ~85% less token intensive for cataloging and classification workloads. Evaluated using OpenEval, even by OpenAI's own tooling, it looked like we had found a wedge.

For a brief window, it seemed like building our own model could become a real moat. But that window didn't last long. By late 2024, we began to realize something deeper - not just about models, but about how value would actually accrue in AI.

Section 01When does it even make sense to build a model?

As we reflected on our journey, one insight became clear: there are only two situations where building or fine-tuning your own LLM truly makes sense. Outside of these two scenarios, building a custom LLM starts to look far less defensible. And retail commerce - the space where we had built our first model - didn't fall into either bucket.

Section 02The harsh reality of cost

Even though our model used dramatically fewer tokens for our task, we were still paying for GPUs at market rates - about $3 per hour for a single unit.

OpenAI and Google don't play that game. They own their infrastructure. Google doesn't even rely on GPUs anymore - they run on TPUs built in-house, at scale. Their capex is already sunk. The marginal cost of an unused TPU in their data center is effectively zero.

For Adya, every hour was real money. Pure OpEx. So even if our model was cheaper per SKU, their unit economics at scale would always beat ours. They could undercut us, bundle it, or simply give it away.

Section 03Distribution is an even bigger moat

Then there was distribution. OpenAI and Google already sit inside every developer workflow, every cloud stack, every enterprise shortlist. Their distribution cost per user is close to zero. For us, every new customer meant sales effort, partnerships, GTM spend. Even if we built something better, the cost of getting it adopted was far higher.

China managed to build strong domestic models because their market was protected. US models couldn't dominate there early. That insulation created space to scale. India doesn't have that luxury. Here, you compete with Google and OpenAI from day one. A few percentage points of better performance won't overcome that gravity.

Section 04The moment of realization: late 2024

By October-December 2024, the landscape shifted dramatically. We saw the arrival of Large Reasoning Models (LRMs) like O1-class systems. Early agentic models - systems that could reason, use tools, browse, plan, and act. Architectures chaining multiple fine-tuned specialists behind a single interface. Mixture-of-experts and reinforcement pipelines that looked simple outside but were deeply complex inside.

This was the turning point. It became clear that what we had built as a fine-tuned commerce model would soon just become another internal expert inside a much larger agentic system. Hyperscalers wouldn't just match our fine-tuning - they would absorb it, chain it, augment it with tools and data, and outperform us again. The moat we thought we had in August was already eroding by November.

That's when we made a conscious decision: Adya would stop chasing model supremacy and start building a platform.

Section 05Why agentic systems change everything

Traditional LLMs optimize for the next token. Agentic systems optimize for outcomes. They reason across steps, call tools, fetch data, verify themselves, and adapt.

So even if a base model doesn't "know" commerce deeply, an agentic system can query catalogs, call APIs, learn from feedback, and outperform a static fine-tuned model over time. Once you combine that with massive infrastructure, continuous data flows, and reinforcement learning, the advantage of a narrow fine-tuned LLM disappears.

Section 06Where we still see enduring value: SLMs

Yet, this didn't mean models no longer mattered. It meant big general LLMs wouldn't be where startups win. The real opportunity lies in Small Language Models (SLMs): models that run inside applications, on edge devices, inside drones, medical devices, or secure on-prem systems, often without GPUs.

Here, the economics flip. External APIs are too costly. Latency matters. Privacy is non-negotiable. And ownership is critical. For defense, healthcare, pharma, banking, and government, this is where intelligence must live. This is also where lock-in still exists.

Section 07The platform bet

So from late 2024 onward, Adya pivoted. Instead of trying to own "the best model," we chose to own the capability to build, tune, orchestrate, and operate models. We built Model Studio - to fine-tune LLMs and create SLMs - alongside tools to embed intelligence into workflows, agents, and edge systems. A platform where enterprises can co-create sovereign models they own, while Adya operates and evolves them.

In this world, the bank owns its credit model. The defense unit owns its edge model. The hospital owns its clinical model. Adya becomes the system of intelligence that designs, runs, and improves them - generating recurring value without fighting hyperscalers on raw inference.

Section 08The lesson from August 2024

Building our first model in August 2024 wasn't a mistake. It was necessary. It taught us how hard real training is, what efficiency gains look like, how quickly the frontier moves, and how unforgiving AI economics can be.

Most importantly, it showed us that the future doesn't belong to those who train the biggest models. It belongs to those who enable intelligence everywhere.

Platforms, not monoliths. Systems, not single models. Outcomes, not tokens. That's the journey Adya chose - and the one we're building for the decade ahead.