Sunday, October 25, 2026

@convequity Oct 12

Convequity on X

Why I think Nvidia buying Reflection AI helps Palantir more than it hurts $NVDA wants its models inside enterprises, not just hyperscalers, neo-clouds and AI labs. Buying Reflection AI would speed that up. The FT reported on 10 Oct that Nvidia is in early talks to buy Reflection, raise its stake, supply it more chips, or hire its team and license its technology. Bloomberg's source says no terms have been reached. Nvidia has already invested $800m, so a buyout or a bigger stake means pricing a company it already partly owns. The enterprise push Nvidia agreed in September to buy Hugging Face for $12.9bn. Jensen Huang says more than 200,000 companies use it to deploy AI. Reflection is validated on Dell's AI Factory with Nvidia, and $PLTR's sovereign AI stack runs on Nvidia's AI factory design. Reflection is built for enterprise code. Its Asimov agent reads a company's code, Slack messages, emails and documents. Its first model, Beam (501bn parameters, 23bn active), is trained for coding. Reflection says the weights ship this month under Apache 2.0, for on-prem or air-gapped use. The bear case: Nemotron is moving too slowly at Palantir One reading is that Nvidia wants Reflection because its Nemotron models are moving too slowly with Palantir. I think that is unlikely, given how Palantir has executed. At AIPCon on 10 Sep, Nvidia's supply-chain team said Nemotron 3.5 Lightning, post-trained in Palantir, beat Nemotron 3 Ultra by a large gap on Nvidia's own benchmark. It runs on-prem and retrains on one GPU in a couple of hours. Cisco will offer Nemotron post-trained with Palantir in its AI Factory with Nvidia. L3Harris said open models fine-tuned on its own data beat the frontier models it had used within 48 hours, at 95% lower cost. Palantir has kept shipping open-model deployments. In June it launched an engine to run Nemotron in sovereign environments, and customers own the retrained models. On 1 Oct Armada became its first certified modular data-centre partner. On 12 Oct $ONDS formed Regna AI on AIP and open-weight models. The likelier case: Nvidia is backing more than one route Palantir fits open models to a company's data, permissions and workflows. Reflection is built to read the company's code. Buying a team that has already built that is quicker than building it in-house. The limit is vendor API rules Reflection reaches a company's systems through the vendors' APIs, and vendors decide what passes through them. $SAP's April API policy bars AI agents that plan or execute API calls, and bulk data extraction, outside SAP-endorsed pathways. Since May 2025, Slack's API terms bar third-party apps from using Slack data to train a large language model or bulk-exporting messages. Those rules apply to Palantir too. In May, SAP made Palantir AIP for data migration an SAP Endorsed App. An independent Reflection would likely need similar approval from each vendor whose systems it reads. Who pays the compute rent Reflection trained Beam on Nvidia GB300s rented from $SPCX at Colossus 2 in Memphis, under a June deal worth up to $6.3bn: about $150m a month through 2029. $NBIS agreed in July to sell it $1bn more capacity. If Nvidia buys Reflection, it takes on those contracts and pays SpaceX rent on Nvidia chips that SpaceX bought. In an acqui-hire, the contracts stay with what remains of Reflection unless the deal says otherwise. The FT says an acqui-hire would likely avoid a long regulatory review. Nebius holders should watch which company ends up owing that $1bn contract. What it means for Palantir The risk is an independent Reflection selling its own forward-deployed team and turnkey stack, which is how Palantir sells. If Nvidia owns Reflection, I expect Beam to reach AIP customers the same way Nemotron does, with each customer fine-tuning it on its own data inside Palantir.