August 22, 2026
New data reported by TechCrunch shows OpenAI closing the gap with Anthropic among business users -- a shift worth pausing on, because it undercuts the assumption that enterprise AI relationships are sticky. Companies are supposedly locked into whichever model they built workflows around. But if usage is swinging back and forth every time a lab ships a new release, that's not loyalty, that's shopping. The article notes this volatility should worry investors on both sides, and I think that's the right read. A subscription business built on switching costs is worth more than one built on 'whoever had the better demo this quarter.' We've made a version of this argument before when Anthropic touted its own run-rate numbers, and it's still true: enterprise AI spend looks real, but which vendor collects it looks awfully unstable. For any business picking a CRM or support stack today, the lesson isn't 'bet on the current leader.' It's build in a way that doesn't assume any one model provider stays on top.
The most useful headline of the day isn't about a bigger model -- it's about a smaller one. Nvidia's research, covered by TechCrunch, found that AI agents can perform reliably even with a mediocre underlying model, as long as the harness around it -- the fine-tuning, guardrails, and scaffolding that structure how the model is used -- is well built. That's a genuinely important finding for anyone shopping for AI tools, because it means the marketing pitch of 'we use the most advanced frontier model' is increasingly beside the point. What determines whether an agent actually does its job without going sideways is the engineering wrapped around it, not the raw horsepower underneath. This is basically the argument for platforms that focus on workflow automation rather than model selection: the orchestration layer is where the reliability actually gets built. It also means smaller companies without frontier-model budgets aren't necessarily at a permanent disadvantage -- a good harness can compensate for a lot. I'd stop short of saying the model doesn't matter at all; there's presumably a floor below which no harness saves you. But Nvidia, of all companies, admitting the model itself is becoming a commodity is a notable moment.
Nvidia's new data center partnership with Cloverleaf and Starcloud's $250 million raise for orbital data centers are two sides of the same story: AI's physical footprint is getting so large that companies are running out of ordinary places to put it, whether that means striking new land deals or looking to orbit. It's a reminder that behind every chatbot update sits an enormous and increasingly strained supply chain. Meanwhile, the DOJ's reported investigation into Andreessen Horowitz -- over partners sitting on boards of competing portfolio companies like Databricks and Fivetran -- is a separate but related signal that regulators are starting to scrutinize the AI money machine more broadly, not just the compute layer. None of this changes what a business should do with AI tools today, but it's worth watching, because if either the infrastructure buildout or the venture-funding structure behind it wobbles, the ripple effects reach every vendor built on top of them.
If model quality is becoming less of a differentiator than the systems built around it, does that change how your team should be evaluating AI vendors right now?
Sources