The AI Brief

Anthropic agreed to a $9.1 billion, 20-year deal with Riot Platforms for 191 MW of AI computing capacity at its Texas data center. Bloomberg reported Monday that the unnamed partner in Riot's SEC filing is Claude maker Anthropic. Riot described the customer in its August 10 SEC filing only as a leading frontier AI lab; Bloomberg subsequently identified the tenant as Anthropic, citing people familiar with the transaction. Neither company publicly confirmed the identity when Bloomberg contacted them.

The agreement runs through June 2048, with two five-year extensions potentially raising total sales to $16.1 billion. Riot expects full deployment by June 2028, delivering the first 96 MW of capacity in December 2027. To fund the buildout, estimated at $2.1 billion to $2.3 billion, Riot secured a $573 million interim financing facility from Morgan Stanley, with plans for 80–90% debt financing. The Rockdale campus has 700 MW of developed, energized power capacity, along with existing fiber and electrical infrastructure that Riot says can be repurposed for high-density computing.

Anthropic has inked several deals with AI cloud computing suppliers in recent months to shore up its computing needs after struggling to keep up with customer demand for AI tools. Riot shares jumped 25% to $24.40 in late trading Monday after the company disclosed the data center agreement.

Disclosure: Claude, which generates this brief, is built by Anthropic.

TSMC disclosed July revenue of NT$467.58 billion ($14.5 billion), up 44.7% year-on-year, reflecting an accelerating appetite for chips used in AI applications. The result puts TSMC ahead of its own full-year guidance. The company expects 2026 revenue to grow by slightly above 40% in U.S. dollar terms, a target it raised after its Q2 earnings, and lifted capital expenditure guidance to between $60 billion and $64 billion for the year.

Because TSMC's customer list spans major AI players such as Nvidia and Google, its monthly revenue releases serve as a broad barometer for technology spending across the industry. The PHLX Semiconductor index has spent six weeks selling off on the theory that AI capex ran ahead of itself, and sits about 15% below its June high. The July results lifted European chip-sector equities, with ASML gaining more than 2% and both Infineon and STMicro moving higher alongside it.

Revenue also increased 5.6% from June, providing a further indication that demand for advanced computing semiconductors remains resilient. TSMC's cumulative revenue for January through July reached NT$2.87 trillion, representing a 37% increase from the corresponding period last year.

OpenAI split its Daybreak cybersecurity program into two access tiers. Daybreak Blue opens frontier general-purpose models, including GPT-5.6 Sol, to approved defenders for everyday security work; Daybreak Red gates the new GPT-5.6-Cyber model behind tighter vetting for vulnerability research, exploit validation, and security testing. GPT-5.6-Cyber answers 95% of sensitive cybersecurity queries in internal benchmarks; standard GPT-5.6 Sol blocks nearly all of them due to its safeguards.

OpenAI used GPT-5.6-Cyber to investigate V8, the JavaScript engine used by Chrome, uncovering two previously unknown vulnerabilities that could be chained to corrupt memory and escape the V8 heap sandbox. Researchers validated the findings and reported them to Google through coordinated disclosure. Google fixed the vulnerability, assigning it CVE-2026-15903, a high-severity flaw where V8's optimizing compiler incorrectly skipped a safety check when converting values to integers. The company also reports at least five vulnerabilities in a widely used mobile operating system, three critical ones in a popular database, and more than 400 privilege-escalation flaws in a popular OS kernel. OpenAI names neither the software nor the vendors, so those claims cannot be checked from outside.

OpenAI's documents list GPT-5.6-Cyber pricing at $12.50 per million input tokens and $75 per million output tokens, compared with Sol at $5 per million input and $30 per million output in the same Daybreak pricing table. Accenture, IBM, CrowdStrike, Cisco, and Palo Alto Networks may put the models into security products and managed services. Entry requires identity verification, account monitoring, approved-use restrictions, and legal attestations. Hardware security keys become mandatory for all Daybreak accounts on September 1, 2026. Under OpenAI's Preparedness Framework, both GPT-5.6 Sol and GPT-5.6-Cyber were assessed as reaching the "High" cybersecurity capability threshold but remaining below "Critical."

Google's Made by Google event is Wednesday, August 12, starting at 3 p.m. PT / 6 p.m. ET. The company will unveil the Pixel 11, Pixel 11 Pro, Pixel 11 Pro Fold, and Pixel Watch 5. Gemini 3.5 Pro has not been officially confirmed for the event, though it has been widely expected, and leaked benchmarks suggest it would compete on visual code generation, spatial reasoning, and agentic tasks at lower per-token cost than Claude Fable 5.

The much-anticipated Gemini 3.5 Pro, initially slated for mid-2026, has faced repeated delays, while the Gemini 3.5 Flash and 3.6 Flash models have failed to meet critical performance benchmarks, leaving DeepMind trailing behind competitors. The model has missed at least five successive public launch targets since Google I/O, and prediction markets earlier priced August 7 at 73% probability, a date that passed without a release. First covered in Vol. I, No. 76.

Google DeepMind's incoming CEO Koray Kavukcuoglu noted in a recent all-hands that Flash is in high demand and Gemma models have surpassed 900 million downloads , but made no commitment on Gemini 3.5 Pro timing. SemiAnalysis published analysis this week concluding that for all intents and purposes, DeepMind is no longer a frontier lab, citing large numbers of departures from reinforcement learning teams and poor compute allocation. Google has not responded to that characterization.

Riot expects full deployment by June 2028, after delivering 96 MW initially in December 2027. The second tranche of 95 MW follows six months later. To fund the buildout, estimated at $2.1 billion to $2.3 billion, Riot secured a $573 million interim financing facility from Morgan Stanley, with plans for 80–90% debt financing. The company has increasingly relied on Bitcoin sales to cover the equity portion: in Q1 2026 alone, Riot sold 3,778 BTC for $289.5 million at an average net price of $76,626 per coin, more than double its quarterly production.

Riot estimates cumulative net operating income of $7.3 billion to $8.2 billion during the base term, though these figures remain forward-looking projections dependent on construction and operational performance. The construction schedule and debt stack mean Anthropic is effectively a pre-paying anchor tenant in a data center that does not yet exist, financed by a company that mines Bitcoin and recently posted a net loss of $237.2 million in Q2 2026.

The deal fits a pattern: leading AI laboratories, including Anthropic and OpenAI, have responded to surging demand by locking in capacity years in advance through multi-year or multi-decade contracts, because traditional cloud providers cannot expand fast enough. As a result, AI companies are turning to alternative sources, sites with robust electrical interconnects, industrial zoning, and proven operational experience with high-load environments. The Riot transaction is the clearest illustration yet of how far that alternative-sourcing logic now extends.

Disclosure: Claude, which generates this brief, is built by Anthropic.