Last week, Salesforce announced Koa, billed in its own press material as the company's "first reasoning model." For anyone who spent the last cycle watching Layer2 rollups promise decentralization while quietly running a single sequencer in a Virginia data center, the shape of this announcement should feel dizzyingly familiar.
Here is the detail that matters, the one every headline buried: Koa is built on Nvidia's Nemotron, an open-source foundation model. Salesforce did not train a frontier model from scratch. It took a public base, ran domain post-training over it, and bound the result to its CRM workflow. That is not a criticism — it is a description. But it is also not the story being sold. The reason this belongs in a crypto column is not that Salesforce touched AI. It is that the exact rhetorical move — take an open resource, wrap it in proprietary framing, and call the wrapper the innovation — has become the defining pathology of our own industry.
To be fair to what Koa actually is, the engineering is real. After the o1 and R1 reasoning paradigm landed, the mainstream recipe crystallized: take a base model, generate chain-of-thought traces, then fine-tune with supervised learning plus reinforcement against verifiable rewards. Koa almost certainly follows that recipe on top of Nemotron. That is combination-level and engineering-level work, not architectural invention, and honest observers should say so plainly.
The base matters because it sets the ceiling. Nemotron is a decoder-only Transformer carrying the standard modern furniture — grouped-query attention, RoPE, SwiGLU — and several of its variants descend from the Llama lineage. Trace Koa's stack back far enough and you arrive at public open-source genealogy, not a proprietary breakthrough. Salesforce's "first" is company-first, never industry-first. Those two words get conflated constantly, and the conflation is the product.
Distribution is the actual asset. Salesforce sits on more than 150,000 enterprise customers, and Koa ships as a capability inside Agentforce and Data Cloud rather than as a standalone API. When your model is embedded in a workflow people already pay for, acquisition cost approaches zero. That is a moat no foundation-model startup can replicate, and it is the only sentence in the entire announcement carrying structural weight.
Now the crypto parallel stops being a metaphor and becomes a diagnosis. For two years, Layer2 teams have described "decentralized sequencing" in pitch decks and shipped centralized sequencers in production. The technology existed; the will to deploy it did not, because a single sequencer is faster, cheaper, and easier to operate. The gap between the roadmap and the mainnet was never a knowledge gap. It was an incentive gap. Decentralization is not a feature you ship; it is a habit you keep — and habits are expensive. Koa's model story lives in exactly that gap.
Consider the economics. Reasoner models consume dramatically more tokens than standard models, because the chain-of-thought is itself output. Under per-consumption billing — and Agentforce pricing already leans that way — the structure pushes cost straight onto the customer. It is the AI equivalent of a gas fee that triples when the network is busy, except here the network is your own sales team. Based on my 2020 workshops teaching newcomers to audit smart contracts by hand, I learned something durable: users do not rebel against cost. They rebel against cost they cannot predict.
The deeper story is commoditization. When application vendors begin building their own reasoning layers, the model itself becomes a commodity and value migrates toward whoever owns the data, the workflow, and the distribution. Nvidia wins twice — once selling the base, once selling the inference stack. Salesforce wins by defending the application. The pure model API vendor is the one left standing in the middle of the road, waving at traffic.
And notice what "built its own model" is really for. In an enterprise market terrified that AI will grind SaaS into paste, owning a model is a trust token and a negotiation lever — it signals you are not merely reselling someone else's intelligence. That is a legitimate move. It should still be read as confidence management aimed at investors, not as a capability claim aimed at benchmarks. The two audiences hear the same sentence and walk away with entirely different beliefs.
The contrarian reading is uncomfortable for my own tribe. We love to dunk on "self-built" claims, and the dunking is easy here. But the honest blind spot is that the decentralization movement holds no monopoly on virtue and no immunity from the same drift. Community is not a user base; it is a shared soul — and yet a vast share of "communities" in this industry are customer segments in costume. We criticize Salesforce for dressing a Nemotron fine-tune as proprietary genius while our own protocols dress a five-key multisig as a DAO. The mirror is unflattering, and it is ours.
The second blind spot concerns data. Koa's real risk is not that it will trail GPT or Claude on general reasoning — it will, and that is acceptable, because CRM tasks reward domain data over raw state-of-the-art. The real risk is that a model reasoning over customer records produces longer visible traces, and longer traces are harder to audit and easier to leak. Enterprise trust is not won by capability. It is won by provable isolation, and that is a design commitment, not a benchmark score. We build not for the token, but for the tribe — and the tribe here is every customer whose records train a system they will never be allowed to inspect.
So watch the signal, not the slogan. If Salesforce publishes a no-training-on-customer-data guarantee backed by third-party audit, Koa matures into a serious vertical benchmark worth studying. If it publishes latency and pricing instead, read the silence carefully. The model was never the moat. The real question is whether the moat protects the customer or the vendor — and that question, as always, is answered in the disclosures nobody prints on the press release.
