The Foundation Nobody Owns
How DeepSeek Became the Substrate the Rest of the Field Builds On
TL;DR: The lab with the best model changes every few months. The lab that becomes the foundation shapes the entire field for years.
The biggest story of this generation of AI was never that the models became smarter. It was that the knowledge stopped staying private.
How DeepSeek Became the Substrate the Rest of the Field Builds On
The reasoning model you used today was very likely trained with an algorithm from a single 2024 DeepSeek paper. The open agent from the previous piece in this series is, by one expert’s description, a supersized version of DeepSeek’s V3. The open model that topped the coding leaderboards this June borrowed DeepSeek’s attention mechanism wholesale and optimized one layer above it. And a model built in India “Sarvam”, trained on government-allocated compute, for twenty-two Indian languages, under none of the export pressure that shaped DeepSeek, arrived independently at the same architecture, down to the routing.
Four labs. Three countries. One floor.
This is the last piece in a series about the Chinese open models, and it is about the lab that poured it. The first three pieces were each about a single model and a single thing it proved: that the cost of intelligence had collapsed Deepseek The New Frontier Is the Invoice, that the cost fell without the quality falling with it GLM: Cheap Was the Easy Part, that even frontier-grade autonomy had gone open and cheap Kimi: The Frontier Went Open. This piece is about what runs underneath all three, and the claim it makes is not the one the leaderboards would suggest. It is not that DeepSeek has the best model. It is that DeepSeek built the ground the others stand on.
Part One: Foundational Is Not the Same as Strongest
It is worth being precise about the claim, because the obvious version of it is wrong.
The obvious version says DeepSeek is the strongest of the Chinese labs. It is not, or at least not reliably, and not this month. On the coding and agentic benchmarks that matter most for real work, GLM-5.2 and Kimi trade the lead between themselves and with the closed frontier; DeepSeek’s flagship is excellent and frequently not first. If the question is “which open model would you reach for today,” the honest answer rotates, and DeepSeek is one name among several.
But “strongest model” is a title with a short half-life. It changes hands every few months, and the lab that holds it in June is rarely the one that holds it in December. There is a different kind of primacy, quieter and far more durable: not the best model, but the methods that everyone else’s models are made of. The strongest model is a fact about one release. The substrate is a fact about the whole field. And on that measure DeepSeek is not one name among several. It is the name underneath the others.
This distinction is the whole piece. Foundational and strongest are different competitions, and DeepSeek decisively leads the one that lasts. What follows is the evidence, three primitives the field now treats as defaults, all of them DeepSeek’s, all of them given away.
Part Two: The Three Primitives
Start with the one most likely to be running in a model you have actually used.
The reinforcement learning method. In 2024, in the DeepSeekMath work, DeepSeek introduced Group Relative Policy Optimization, GRPO. The technique strips the expensive critic model out of reinforcement learning, estimating its baseline from the spread of a group of sampled answers instead of training a second network the size of the first. It was, in the plainest terms, a way to do RL far more cheaply, and it became one of the main drivers of R1, the model that made reasoning training famous. It did not stay DeepSeek’s. GRPO is now a default across the field, the method reached for when a lab wants reasoning behavior without the memory overhead the old approach demanded, used well beyond China and well beyond the open models. A single paper changed how the industry does reinforcement learning.
The memory mechanism. DeepSeek-V2 introduced Multi-Head Latent Attention “MLA”, which compresses the Key-Value cache, a model’s working memory, into a small latent representation rather than storing it in full, and does so without the accuracy loss that cheaper compression tricks incur. This is the mechanism, described in the previous piece, that makes long-context and long-horizon work affordable. Kimi’s trillion-parameter agent runs on it; one reviewer’s description of that model as a supersized DeepSeek-V3 is a description of MLA scaled up. The attention that lets an open agent hold hundreds of tool calls in view is DeepSeek’s design, adopted and enlarged.
The architecture recipe. DeepSeekMoE, fine-grained experts, a shared expert held always-on for the common case, is the Mixture-of-Experts blueprint the field converged on. Qwen’s 235-billion-parameter MoE has been described by architecture analysts as very similar to DeepSeek-V3’s. And the most telling evidence is the Indian model, Sarvam, built under no export constraint at all: it adopted the same playbook, the small fine-grained experts, the shared-expert routing, not because anyone forced it to, but because, by its builders’ own account, these had simply become the right answers for building efficiently at scale. When a design propagates to a lab that had no reason to copy it, the design has stopped being a choice and become a standard.
Three primitives. The way you train reasoning, the way you store memory, the way you structure the model itself. Each was born of constraint, invented because DeepSeek could not buy its way around a limit and had to engineer through it, and each was published the moment it worked. The constraint produced the methods; the openness spread them.
Part Three: The Pattern Is Still Running
None of this is history. The same thing is happening in 2026, in real time, with the newest work.
The coding model that topped the leaderboards in June, GLM-5.2, did not invent its own attention. It integrated DeepSeek Sparse Attention “DSA”, DeepSeek’s more recent design, and built its headline efficiency feature one layer above it, as the GLM piece in this series documented in detail. Inheritance, again, with optimization on top.
And in late June, DeepSeek published something that sharpens the whole argument. Alongside the DSpark speculative-decoding system, it open-sourced DeepSpec, not a model but the training framework itself, MIT-licensed, the machinery for building and evaluating the draft models that make inference faster. DeepSpec ships several algorithms together, and DeepSeek demonstrated it working not only on its own models but on other labs’, on Gemma, on Qwen. This is the foundational move in its purest form: a lab releasing not just the output of its research but the apparatus to reproduce and extend it, designed from the start to run on everyone’s models. DeepSpec is two weeks old as of this writing; nothing has inherited it yet. But if the pattern of the last three years holds, something will, and soon. The foundation is not a finished thing. It is being poured continuously.
Part Four: A Commons, Not a Monoculture
A skeptic should object here, and the objection is the most important part of the case.
If everyone is building on DeepSeek’s primitives, is this not just imitation, a field of followers copying one leader? It is not, and the proof is that the field also rejects DeepSeek’s ideas when they fail to hold. MiniMax built a model on linear attention, an efficiency bet adjacent to the sparse-attention family, and then reversed it in its next release, stating plainly that linear attention proved tricky in production, accurate enough for ordinary prompts but weak on the multi-turn reasoning that real work demands. Qwen, in its latest series, dropped the shared expert that DeepSeek’s recipe holds central. These are not the moves of labs copying a template. They are the moves of labs testing one, adopting what survives contact with their own workloads and discarding what does not, in public, failures included.
That is the difference between a monoculture and a commons. A monoculture copies. A commons experiments on shared ground and reports back, so that the next lab inherits not just the technique but the knowledge of where it breaks. The reversals are not evidence against the foundation. They are evidence that the field is doing genuine research on top of it.
And there is a deeper turn the honest version of this argument has to make: the foundation is itself founded on what came before. DeepSeek did not invent the Mixture of Experts, or the Transformer, or reinforcement learning; GRPO refined a lineage of policy-optimization methods, and DeepSpec, the framework just discussed, is built openly on an earlier open project. DeepSeek poured a slab, but it poured it on ground that others had already leveled, using tools others had already made. Which is precisely why the title of this piece is what it is. The foundation has no owner. Not even the lab that poured most of it owns it, because the moment each piece worked, DeepSeek gave it away, and because DeepSeek, too, was standing on a commons when it built.
Part Five: The Anti-Enclosure
There is an economic shape under all of this, and it is the inverse of the one that runs through much of the rest of this newsletter.
The recurring pattern in the enclosure economy is the fencing-off of something once freely available, so that a metered replacement can be sold in its place, value created by withholding. The Western proprietary frontier runs on a version of this logic, and runs on it rationally: weights held private, capabilities exposed only through a metered interface, each advance kept rather than shared because the advantage is the keeping. In that world an innovation is a moat, and a moat works only as long as no one else can cross it.
The open field operates on the exact opposite principle, and the inversion is not idealism, it is mechanics. When DeepSeek publishes GRPO, GRPO does not lose value by being shared; it becomes the baseline the whole field starts from, and the field’s progress accelerates, and DeepSeek sits at the center of an ecosystem it seeded. A moat divides. A foundation compounds. This is the real answer to the question the first three pieces kept raising but never fully resolved, why the constrained regime closes the gap to the frontier as fast as it does. It is not only that export controls forced the efficiency. It is that openness let the efficiency propagate. The wall produced the methods; the commons multiplied them. An enclosed frontier can only move as fast as one lab. An open one moves as fast as all of them at once.
The Lab Underneath
The series opened on an invoice, DeepSeek’s $0.87, the price that revealed cost had become the architecture. It closes on something the invoice only hinted at. The cheap price was never the most important thing DeepSeek produced. The most important thing was the set of methods it could not stop others from using the moment it showed them working.
The strongest open model will keep changing hands. It may be GLM this quarter and Kimi the next and a model not yet released the quarter after. But underneath whichever one wins sits the same floor, the reinforcement learning method, the attention mechanism, the architecture recipe, and now the training frameworks, and one 160-person lab in Hangzhou poured most of it, then handed it to everyone, then went back to pouring more. That is a stranger and more durable kind of power than a leaderboard position. The lab that owns the best model owns a moment. The lab that becomes the foundation owns the shape of what everyone builds next, and owns it, fittingly, by not owning it at all.
The most powerful model wins a quarter. The methods everyone builds on shape a decade.
The series:
Sources
A moment of history comes to mind. The moment serious compute moved onto an office desk in a way usefull enough to pay the very serious price for the hardware. Till then they had been a very expensive and exciting hobby for tech heads. Huge corporate mainframes and timeshare access were the only real game in town.
Visicalc. Business or planning data laid out on your new screen, recalculating numbers live as you shifted assumtions. All in the form completely familiar, a standard tool on paper. But on paper they were hard work for planning or testing ideas. Way too many hours of calculating for the workflow that the completely new device on your desk allowed.
It may be that DeepSeek's most radical contribution is the power of local LLMs. That changes everything. In a way that is rather parallel to the Visicalc moment. A not expensive SBC or the Mate 50 can run a decent LLM already, as the core interface and orchestration of the s]stem. That is just the start. What can be done with that is just emerging, like what could be done with a spreadsheet once it was no longer graphite put onto paper manually. Mainframes and timeshare did not go away, and timeshare returned as cloud. Local compute shifted the user base and the economics.