Salesforce built Koa, its own reasoning model, on Nvidia's open weights

Salesforce's Agentforce used to route any multi-step reasoning job out to a frontier model. "But reasoning has always been something that we've relied on the frontier model providers for. Until now," Jayesh Govindarajan, EVP of Salesforce AI, told TechCrunch ^[3]. What replaced that route is Koa, post-trained from NVIDIA Nemotron 3 Super with supervised fine-tuning and reinforcement learning on NeMo RL, NeMo Gym and NeMo AutoModel, on a synthetic dataset drawn from close to three decades of CRM deployments across more than 14 industries ^[1][4]. Because Nemotron is open, Salesforce fine-tuned and runs the model inside its own infrastructure and holds the weights ^[5]. Salesforce says Koa matches or exceeds leading model performance on CRM actions with three times fewer errors on its own CRM Bench ^[6].

Rohan Kumar, Salesforce's president of platform and engineering, said of the training data: "Not a single byte of customer data was used" ^[7].

So the model-layer capability was assembled without a pre-training run, without customer records, and without the frontier lab. What Salesforce could not assemble was the base model, and Govindarajan was specific about why Nemotron was the one: "the challenge has always been the lack of a pre-trained base model to start with. Until Nemotron came along, there was no sovereign American pre-trained model that was available, one, and two, that was state of the art, and, three, that had clear data provenance. We have no idea what Qwen trains on" ^[8].

The distribution question sits one layer down

That is a dependency. Jensen Huang, on the Dreamforce stage with Marc Benioff, put open-model adoption at "about 30% at the beginning of last year to some 70% now" ^[9][10]. That number is the vendor's own. NVIDIA supplies the open base, the post-training toolchain and the inference architecture that Kari Ann Briski, NVIDIA's VP of Generative AI Software for Enterprise, described as "a unique architecture for inference to be token efficient" ^[11]. Salesforce did not leave the frontier labs either; it announced ClaudeForce alongside Koa, letting customers use Claude as the interface while data stays in Salesforce's system of records ^[12].

The same week, Nvidia was on the other side of that trade. Nvidia confines Anthropic's Claude Fable to lower-stakes work such as open-source projects and uses Nemotron for sensitive jobs like supply chain monitoring, The Information reported ^[13]. "As a company, you know, we believe ZDR [zero data retention] should be on by default," said Justin Boitano, Nvidia's vice president of enterprise AI ^[14]. Nvidia is an Anthropic investor and Anthropic is a major buyer of NVIDIA hardware ^[15].

What the retention clause cost

The trigger was a June Anthropic policy giving it the right to keep customer usage logs for 30 days, and up to two years when its safety systems flag them, covering Claude Fable 5 and removing a zero-data-retention guarantee some enterprise customers had relied on ^[16]. Booz Allen barred employees from using Fable on work touching the proprietary cybersecurity software it sells, with CTO Bill Vass saying, "We worry a little bit that [Fable] might be learning from some of our code," while noting the concern covers only a small share of the firm's work ^[17]. Northrop Grumman runs open-source models on its own air-gapped servers; Novo Nordisk keeps proprietary data out of hosted models ^[18]. Microsoft restricted its own employees' access while its legal team reviewed the change, PYMNTS reported at the time ^[19].

Anthropic answered on September 1 with Enterprise Frontier Safeguards: retained data sits in the customer's own cloud under the customer's encryption keys, automated safety monitoring continues, and the customer's own security team reviews what gets flagged, with broader availability targeted for this fall ^[20]. Palantir will not offer Fable through its platform until Anthropic provides an irrevocable zero-data-retention guarantee, though Palantir customers can still buy the model directly from Anthropic ^[21]. An executive at a major U.S. utility told The Information the company scrapped plans to test Fable on core power infrastructure serving millions of households after Anthropic would not give the same assurance, and still uses Anthropic for finance and HR work ^[22].

This was a pricing event. The utility kept Anthropic for lower-stakes work, so what it balked at was the clause, and what Palantir is withholding is a channel.

Two engineers, two months, 40,000 lines

The build-side evidence points the same way. Perplexity was paying too much for DynamoDB and getting too little control over read performance. So it replaced DynamoDB reads in its search serving layer with CobbleDB, a roughly 40,000-line Rust key-value store written by two engineers in two months with help from hundreds of persistent AI coding agents, The New Stack reported ^[23]. At roughly 200,000 requests per second, Perplexity measured median batch-read latency of 5.6 milliseconds against 31.4 ms previously recorded on DynamoDB, and p99 of 24.2 ms against 123 ms ^[24]. The report is explicit that the systems were not tested side by side on identical traffic, since the DynamoDB figures predate the cutover and CobbleDB's come after ^[25]. Later load testing reached 500,000 requests per second before performance declined ^[26].

The median improvement is 31.4 to 5.6 ms, a reduction of 25.8 ms, or about 82% ^[27]. The p99 improvement is 123 to 24.2 ms, about 80% ^[28]. Both figures come with the caveat above, so they are directional. The projected saving is at least 20% against DynamoDB, and the estimate excludes the engineering time to maintain the thing ^[29]. Weighing a saving of at least 20% on a database bill against two engineers' ongoing time is a calculation only Perplexity can do, and the report does not include the DynamoDB bill.

The two engineers kept control of the architecture and the production system, and Perplexity kept the agents out of production ^[30]. Andy Pavlo of Carnegie Mellon said at Percona Live this year that databases are the hardest and most important challenge for AI agents, partly because mistakes involving production data can be difficult or impossible to reverse ^[31].

Where defensibility actually sat this week

Set those two against Zocdoc, which is selling nothing model-shaped at all. Zocdoc is opening the scheduling infrastructure behind its own site to Yelp, Healthgrades and Amazon Health AI, with separate financial arrangements for each partner, Fast Company reported ^[32]. In the Blue Shield of California pilot, Zocdoc's tools were added to a provider directory that previously listed doctors without letting anyone book. The insurer says the directory now surfaces more than two million hours of appointments across a 90-day window, roughly double where it started, with booked appointments up sevenfold, an average wait of six days and a no-show rate of 2% ^[33]. The national average time to get an appointment with a physician hovers around 31 days, per MedPage Today ^[34]. Six against about 31 days is a reduction of roughly 25 days, about 81% ^[35].

That asset took a decade and change to build. The network covers 200,000 providers, 20 million patients and 10,000 insurance plans, growing at a 50% compound annual rate since 2023, and getting there meant connecting to thousands of separate scheduling systems ^[36]. "Healthcare has a resistance to disruption," founder and CEO Oliver Kharraz said. "It's the brownfield industry where $5 trillion has already been poured into infrastructure that you cannot easily innovate around" ^[37]. Khosla Ventures led the $3 million Series A in 2008, joined the $15 million Series B in 2010, and in May offered to buy some stock at $21 per share, valuing Zocdoc at $2.3 billion ^[38]. A Khosla managing director familiar with the investment called it "patient capital" ^[39]. Eighteen years from a $3 million round to a $2.3 billion mark is the price of the integration work ^[40].

And the distribution is not free even for the company that owns it. Kharraz hired Brian Irving in 2026 as Zocdoc's first CMO because, he said, "When we poll people why they don't use Zocdoc, the number one reason by a large margin is that they didn't know it existed" ^[41].

The counter-thesis, which is not weak

I think the week's evidence points to process and channel as the scarce input, and I'd expect enterprise AI margins to accumulate wherever a retention clause or an integration is hard to replicate. Two ways that reading fails. First, the Koa case is a dependency dressed as independence: Salesforce needed an open base model with clear provenance and there was one, from a supplier with its own reasons to seed the category ^[8]. And Salesforce's performance figures are its own, on its own benchmark ^[6]. Second, Perplexity's cost claim is a projection that leaves out maintenance labour and rests on latency numbers the report says were never measured side by side ^[25][29]. If either the base-model supply or the agent-built-infrastructure economics turn out worse than they look this week, the model layer is expensive again and the argument goes with it.

The falsifier for the channel thesis is narrower. If Anthropic's Enterprise Frontier Safeguards ships broadly this fall and Palantir drops the irrevocable-ZDR condition, then the terms were a temporary negotiation ^[20][21]. I would watch the platform listing itself.