Editor's note. Today leads with money following reliability, not raw capability: Scaled Cognition's fresh $100M round is a bet that the next enterprise AI fight is about whether a model can be trusted to run in a regulated workflow at all. The second story — Samsung re-admitting ChatGPT after a three-year ban, but walling off its chip unit — shows what that trust actually costs in practice. The third is the open-weight counter that keeps getting harder to argue with on coding. Regulation appears only in the quick hits, so there are zero regulation mains today.
The enterprise AI bet shifts from "most capable" to "reliable enough to deploy"
On 25 June, Scaled Cognition raised $100 million in a Series A led by Khosla Ventures, with Genesys participating, at a roughly $750 million valuation. The pitch is narrow and pointed: not a more capable frontier model, but a more reliable one. The company is commercialising a model it calls APT — an "Agentic Pretrained Transformer" — purpose-built for high-stakes customer interactions where a hallucinated answer carries real consequences, wrapped in a platform for simulation, monitoring and evaluation. Founders Dan Klein (a Berkeley AI professor) and Dan Roth previously built and sold an early agentic-AI company to Microsoft, and the startup says it is already in production with Fortune 500 firms across financial services, healthcare, telecom and insurance, on track to automate over a billion customer interactions in the next year.
The interesting signal isn't the cheque, it's where the frontier of enterprise value is moving. For two years the competition was a benchmark race; this round prices the opposite thesis — that in regulated sectors the binding constraint is trust, not peak capability, and that the deployable model is the one whose failure modes you can characterise and contain. That is the right diagnosis, and it is also incomplete for a European buyer. Reliability is necessary but not sufficient when the data is regulated: a model you can trust to answer correctly but cannot run inside your own jurisdiction still fails the compliance test before the accuracy test is even reached. The lesson Stav keeps returning to is that "deployable" in regulated Europe has two axes — behavioural reliability and jurisdictional control — and a frontier that optimises only the first leaves the second for you to solve.
Samsung re-admits ChatGPT — and shows what the price of admission is
The same week, Samsung Electronics deployed ChatGPT Enterprise and Codex to all employees in South Korea and across its global Device eXperience division — one of OpenAI's largest enterprise rollouts to date, spanning the teams behind its phones, TVs and consumer electronics. What makes it more than an adoption headline is the history: this same company banned ChatGPT in 2023 after employees leaked source code into it. The model didn't change Samsung's mind; the wrapper did. It is back only inside ChatGPT Enterprise's governance layer — data-protection guarantees, access management and security controls that fit the model into Samsung's own policy framework — and Codex weekly users in Korea have since grown nearly 800%.
The detail worth holding onto is the part that stayed banned. Samsung's semiconductor Device Solutions unit remains under tighter restrictions — the crown-jewel data still doesn't touch the hosted US model. Read together, those two facts are the whole enterprise-AI playbook in miniature: governance controls are the price of admission for a frontier model from another jurisdiction, and for your most sensitive workloads even that price is too high, so you carve them out entirely. That is the same calculus an EU bank or hospital runs, except its "Device Solutions unit" is most of its book of business. The governance wrapper makes the everyday case workable; it does not solve the sovereign case, which is precisely the gap a model you host yourself is meant to close.
The open-weight counter crosses the closed line on coding
Against both of those — the rented frontier and its governance tax — the self-hostable option keeps gaining. Independent write-ups this month confirm Z.ai's MIT-licensed GLM-5.2 beating GPT-5.5 on SWE-bench Pro (62.1 to 58.6) and landing within a point of Claude Opus 4.8 on the long-horizon FrontierSWE benchmark (74.4 to 75.4), at roughly one-sixth the per-token cost. It is a 744-billion-parameter mixture-of-experts model with a million-token context window, downloadable from Hugging Face and runnable on infrastructure you control.
The structural point is the one this brief has tracked all spring, now sharper: open weights stopped being the budget compromise and became a credible default at the top of a real, multi-step coding benchmark. For a regulated European buyer that matters because it removes the last comfortable excuse for accepting a governance wrapper around someone else's hosted model — "the open option isn't good enough yet" is no longer true on coding. The standing caveat still holds: these are vendor and third-party numbers, and a SWE-bench Pro score is not your production workload, so the honest move is to reproduce the result on your own data before committing a regulated system to it. But the direction of travel is unambiguous, and it points at a stack you own.
Quick Hits
- Gartner puts 2026 AI-agent software spend at $206.5B — up 139% from $86.4B in 2025 and forecast to reach $376.3B in 2027, with the firm expecting 40% of enterprise apps to embed task-specific agents by year-end, up from under 5% at the start of it. The deployment phase has arrived; the reliability and jurisdiction questions above are now budget-line problems, not lab problems.
- The EU AI Act Omnibus reaches its Council vote on 29 June — after Parliament adopted the text on 16 June, the Council is expected to formally adopt it on Monday, with Official Journal publication to follow. Standalone Annex III high-risk obligations move to 2 December 2027 and the watermarking grace period to 2 December 2026; the original text remains law until OJ publication. Three days out, this closes a thread that has run for weeks.
- NVIDIA shipped Cosmos 3 as an open physical-AI omnimodel — released on Hugging Face in two sizes, a Nano (8B reasoner + 8B generator) and a Super (32B + 32B), combining vision reasoning, world generation and action prediction for robotics and synthetic data. Open weights now extend past language into the physical-AI stack — though "open" here comes with the usual licensing caveats worth reading before you build on it.
- Claude Fable 5 stays dark — two weeks after the 12 June US export-control order, the model and Mythos 5 remain offline with no restoration date, and Anthropic's identity-verification path (government ID plus facial geometry — per CIO's reporting, potentially a US-consumers-only route around the ban) takes effect 8 July ahead of the 1 August executive-order deadline. The episode's standing lesson is unchanged: access to a hosted frontier model is governed by someone else's compliance machinery.