October 20, 2026
TechCrunch reported this week that the US government has ordered federal civilian agencies to strip DeepSeek, and any third-party tools built on its models, out of their systems entirely. The official reason is national security, and that's not a surprising move -- governments have been nervous about where their data ends up for years, and a Chinese-built model with free and wildly capable weights was always going to draw scrutiny once it got good enough to matter. What's notable here is the scope: this isn't just 'don't use the DeepSeek app,' it's 'rip it out of anything built on top of it,' which tells you how deep DeepSeek's cheap, open models have already spread into unofficial tooling across government IT. For business readers, the lesson isn't about DeepSeek specifically -- it's that model provenance is quietly becoming a procurement requirement, not just a technical footnote. If you're building internal tools on top of any AI model, you now need to ask not just 'is it good and cheap' but 'will this still be allowed to touch our data in twelve months.' That's exactly the kind of risk conversation we have with customers evaluating a no-code app builder versus stitching together their own AI stack -- the underlying model matters, and so does who's accountable for it, which is part of why we publish a security and trust center rather than leaving customers to guess.
On the other side of the ledger, TechCrunch also reported that Nscale and Anthropic are in talks for a compute deal reportedly worth $45 billion -- yet another entry in what's become a genuinely staggering string of infrastructure commitments from Anthropic over the past year. At some point the number stops being newsworthy on its own and becomes a signal about the industry's cost structure. Every frontier lab is locking in multi-year, multi-billion-dollar compute commitments because none of them believe demand is going to slow down, and because whoever runs out of GPU capacity first loses customers to whoever didn't. My honest take: this is good news if you're a enterprise buyer who wants frontier-quality models to keep getting cheaper per token over time, but it's a warning sign if you're building a business that depends entirely on one lab's API pricing staying stable. These deals get renegotiated, labs shift priorities, and the companies paying the bill eventually pass costs downstream. We've written before about how fast loyalty between AI labs and their biggest customers can shift -- see our take on Anthropic, Replit, and the myth of AI loyalty -- and nothing in this new deal talk changes that calculus.
The third piece of the compute puzzle: Amazon's new TorchNeuron dev tool, which TechCrunch described as Amazon's biggest attempt yet to break Nvidia's grip on PyTorch, the toolkit most AI developers use to build their apps. If Amazon can get even a meaningful slice of developers building against its own chips instead of defaulting to Nvidia hardware, that's a real dent in Nvidia's pricing power -- and over time, cheaper, more competitive inference for everyone buying AI capacity. I'm skeptical it moves fast; developer habits around PyTorch and CUDA are sticky for a reason, and Nvidia didn't build its moat by accident. But it's worth watching alongside the Anthropic-Nscale compute arms race above, because the two stories are really the same story: the fight over who controls AI infrastructure is heating up on every front, from chips to clouds to raw dollars committed. Businesses evaluating AI vendors should treat 'which compute stack are they locked into' as a real due-diligence question, not an afterthought.
Put these three stories together and a pattern emerges: the AI industry is consolidating around a handful of trusted, heavily-capitalized infrastructure bets, while simultaneously drawing harder lines around which models and vendors are considered safe to use at all. That's a reasonable response to real risk, but it also means the cost and complexity of 'just picking an AI vendor' keeps climbing for ordinary businesses that don't have a procurement team parsing export-control news every morning. It's part of why we think the smarter move for most small and mid-size teams is choosing a platform that abstracts that churn away rather than betting on any single model or chip vendor directly.
If your company had to audit every AI tool in use tomorrow for where its underlying model comes from, would you actually know the full list?
Sources