The compliance era comes with a watermark | Wednesday, August 12, 2026

Anthropic struck a roughly $9.1 billion agreement with Riot Platforms, a Bitcoin miner now selling AI data-center capacity, for about 191 megawatts out of Riot's Rockdale, Texas site, phased in through 2027 and 2028. It is another sign of how far labs will reach for power: a crypto miner's shell becomes frontier-AI infrastructure because it already has the grid connection. For buyers, the read is supply, not novelty. Your vendor's capacity increasingly sits in repurposed, hastily contracted sites, which affects how durable and how cheap it stays. Disclosure: Anthropic is a Spearhead technology partner, and the harder read is the one we ran.

OpenAI's public IPO prospectus is expected on SEC EDGAR in the next couple of weeks, with a September listing still the target and a valuation discussed near a trillion dollars. The financials it will expose are the point: roughly $2 billion in revenue a month and more than $20 billion in annual recurring revenue, against a projected $14 billion loss this year and no profit expected until 2029. The most important AI vendor to enterprises is about to open its books, and what buyers will see is enormous demand funded by enormous losses. Pricing power, and your future rates, depend on which one wins.

OpenAI said on August 7 it slowed work on Astra after internal tests could not rule out that the model reached the "critical" cyber threshold in its Preparedness Framework, meaning the ability to find and exploit vulnerabilities in hardened systems without human help. It is the first time a lab has publicly held back a model on cyber grounds, and OpenAI plans to have government agencies and outside safety groups validate the model before release. For enterprises, this is the pre-release-review era becoming concrete: the most capable models will arrive later, tested harder, and with a paper trail you can point your own auditors to.

Mistral rolled out regional inference endpoints that let customers pick whether workloads run in Europe or the US, a priority tier with an uptime guarantee, and European Compute Units that convert multi-year cash commitments from Amadeus, ASML, and CMA CGM into 200 megawatts of infrastructure by 2027 and a gigawatt by 2030. It will also host third-party open models, starting with Z.ai's GLM-5.2. For a regulated European enterprise weighing where its data runs, this makes data residency a checkbox, not a project, and it gives buyers a credible non-US frontier option with committed capacity behind it.

ServiceNow expanded its AI Control Tower to discover, observe, govern, secure, and measure any AI system across an enterprise, not just AI built on ServiceNow. The pitch is no longer build an agent, it is keep track of the dozens you already have. As agents spread across SaaS tools, the hard problem shifts from deployment to oversight: knowing what is running, what it can touch, and whether it is compliant. For a CIO, this is the governance layer that the watermarking rules and the AI Act are about to make mandatory, sold as a dashboard. The category to watch is AI that governs other AI.