Nvidia is spending roughly $26 billion subsidizing open-source AI model development, according to a Wired report cited by AI researcher Nathan Lambert, who helped build the open-source Olmo models at the Allen Institute for AI. Meta, separately, keeps releasing its own frontier models as free weights with no revenue attached. Lambert’s argument, laid out in a new Interconnects post, is that both companies are giving away intelligence for structurally different reasons, and the distinction determines which strategy has staying power.
Nvidia’s bet runs through what Lambert calls the open-source recipe: full training data, code, and methodology, the model Olmo and EleutherAI’s earlier Pythia project represent. Anyone with enough compute can take that recipe, retrain it, and ship new weights. Nvidia’s own Nemotron models follow the same pattern, releasing data and training code wherever licensing allows. The point, in Lambert’s telling, is to stop intelligence from consolidating inside two or three labs so that demand for inference, and therefore for Nvidia’s chips, spreads across as many builders as possible.
Meta’s incentive runs the other way. Releasing a strong model like Muse Spark 1.2 as open weights cuts directly into the token-selling businesses of Anthropic and OpenAI. Meta does not need model access to generate revenue; advertising already funds its balance sheet. Giving away a competitive model for free lets it suppress the pricing power of labs that depend on API revenue, without building any inference-demand loop of its own.
Both moves fit the old software-industry logic of commoditizing a complement, Lambert argues, but the mechanics diverge. Nvidia needs the open ecosystem to become self-sustaining, generating enough inference demand within a few years to justify what it has spent. Meta needs no such loop. It can keep flooding the field with capable free weights indefinitely, bankrolled by a business that has nothing to do with token sales.
Lambert’s skepticism cuts toward Nvidia’s side of the bet, not Meta’s. He notes that open-source labs largely have not bowed out under the cost pressure his framework predicts: Databricks and 01.ai stepping back from frontier training reads to him as an anomaly rather than the start of a trend. That is thin evidence either way. It shows the recipe approach still has runway, not that it is working.
We reported yesterday that Anthropic’s chief executive argues open weights mainly shift power toward whoever owns the most compute rather than democratizing access. Lambert’s account complicates that framing by treating Nvidia and Meta as running two distinct plays with different endpoints, not one uniform transfer of power.
What would undercut Lambert’s case is straightforward: if the list of labs abandoning frontier open-source training grows past Databricks and 01.ai over the next few years, or if inference demand does not scale with Nvidia’s subsidy, his fallback scenario takes over. Open models stop chasing general frontier capability and specialize instead, becoming a long-tail of enterprise agents fine-tuned for narrow, on-premises tasks rather than a true alternative to closed labs.
For operators building on open-weight models, the practical signal to track over the next two quarters is whether more labs join Databricks and 01.ai in stepping back from base-model training. A stable or shrinking list keeps Nvidia’s subsidy thesis alive; a growing one argues for treating open weights as a specialization tool rather than a bet on frontier parity.
Adapted from Nathan Lambert’s analysis in Interconnects, published August 17, 2026.
