Four of the five things that crossed my feed today are promises with a delivery date stapled to them, and the one that actually shipped got no coverage at all. Here is the operator’s read on which of these you can build against this morning and which one is a press release wearing a product’s clothes.<\/p>\n\n\n\n
Bloomberg broke the story Tuesday that Anthropic has committed to a six-year, $10 billion compute deal with Volta, and TechCrunch’s writeup<\/a> fills in who Volta is: an AI cloud startup founded earlier this year by former Brookfield managers, sitting inside Nvidia’s cloud partner program. The capacity itself is Nvidia Vera Rubin silicon in a Norwegian data center that is not built.<\/p>\n\n\n\n The real numbers are in Bitdeer’s own announcement of the lease<\/a>, because Bitdeer is a public company and had to file. Its subsidiary signed a 16-year colocation and services agreement covering 121 IT megawatts at the Tydal campus, worth roughly $4.7 billion over the base term, with an eight-year renewal option that would take it to about $8.0 billion across 24 years. The rate is roughly $202 per kilowatt-month with 3% annual escalators, at a power usage effectiveness around 1.1 on entirely renewable supply. Phase one targets December 31 of this year. Phase two targets March 31, 2027.<\/p>\n\n\n\n Worth noticing what Bitdeer does not say: it never names Anthropic, only “a leading AI lab.” The customer identification is Bloomberg’s, sourced to people who were not authorized to speak.<\/p>\n\n\n\n Here is the number I would actually stare at, and it is none of the ones in the headlines. Roughly $1.3 billion in letters of credit is anticipated from J.P. Morgan affiliates and one other global institution. A company founded six months ago cannot sign a 16-year lease against a crypto miner’s campus on its own balance sheet, so a bank is renting its credit rating into the middle of the transaction. That is precisely how commercial real estate has always been financed. It is now how frontier compute gets financed, which means the counterparty question for anyone building on Claude in 2027 is not “does Anthropic have capacity” but “did phase one energize on time.” Bitcoin miners, a six-month-old lessor, and a letter of credit are all load-bearing in that answer now.<\/p>\n\n\n\n Verdict: matters.<\/strong> The supply you will be metered against in 2027 is being poured in concrete this quarter.<\/p>\n\n\n\n onsemi does not make accelerators. It makes the power tree underneath them, which is why its quarter reads as a cleaner proxy for capex than any hyperscaler’s. The company’s second-quarter release<\/a> puts revenue for the period ended July 3 at $1,604 million, up 9% year over year, with non-GAAP gross margin of 39.3%, non-GAAP diluted earnings of $0.74 a share, and free cash flow of $425.4 million, quadrupled from the year-ago quarter. Chief executive Hassane El-Khoury said revenue, margin and earnings all came in above the midpoint of guidance, and that “AI data center remains our fastest-growing business, and we now expect revenue to more than double in 2026.” September-quarter guidance is $1,650 million to $1,750 million at 40.0% to 42.0% non-GAAP gross margin.<\/p>\n\n\n\n Verdict: matters, with a footnote.<\/strong> Quadrupled free cash flow at the picks-and-shovels layer beneath the picks and shovels is not a narrative, it is a purchase order that already cleared. Two cautions though. “More than double in 2026” is a forecast about a year with two quarters still to run, and plenty of yesterday’s coverage repeated it as a completed fact. And the AI wins the release actually names are an expanded role in the Nvidia MGX ecosystem plus a platform win with Great Wall, a China cloud infrastructure power supplier, not the American hyperscaler roster some of the secondary writeups implied. Read the filing, not the summary of the filing.<\/p>\n\n\n\n The August 4 entry in Google Cloud’s Gemini Enterprise release notes<\/a> adds end-to-end tracing for data connectors, with two new spans: I have burned more hours reconstructing why an agent called the wrong tool than I have writing agents. Two trace spans will never be a headline, but Verdict: matters.<\/strong> It is the only thing on this list you can use before lunch.<\/p>\n\n\n\n Alibaba announced Qwen3.8-Max on Monday, a 2.4-trillion-parameter mixture-of-experts model with a million-token context ceiling. The Qwen team made it broadly available through the API and said the open weights ship next week, which MarkTechPost’s technical breakdown<\/a> confirms is still a promise rather than a download.<\/p>\n\n\n\n I counted three separate outlets running some version of “Alibaba unveils open-source Qwen3.8-Max” within a day of the announcement. As of this morning it is a closed API model with a calendar promise attached, no named license, and no disclosed activated-parameter count. It therefore cannot answer the only question that decides anything for an operator: can I run this inside my own environment, and under what terms. Open weights without a license file is a marketing category, not a deployment option.<\/p>\n\n\n\n Two details the fast coverage skipped, both of which cut the same direction. Alibaba’s multimodal comparison table benchmarks against Qwen3.7-Plus rather than Qwen3.7-Max, which flatters the generational jump. And that million-token window is a ceiling, not a working budget: practical input tops out near 991,000 tokens, or about 983,000 with thinking enabled.<\/p>\n\n\n\n Verdict: marketing, until the weights land.<\/strong> If they arrive next week under a permissive license, it is the most consequential release of the month and I will say so in this slot. None of that is true today.<\/p>\n\n\n\nON Semiconductor Says Its AI Data Center Revenue Will More Than Double<\/h2>\n\n\n\n
Google Shipped Two Trace Spans and Nobody Wrote It Up<\/h2>\n\n\n\n
execute_tool<\/code> and invoke_connector<\/code>. A day earlier, seventeen new data stores went to public preview, Oracle NetSuite, S&P Global, Moody’s, Attio and Supabase among them. On August 1, the pay-as-you-go edition went generally available, billing on consumption instead of pooled user license quotas.<\/p>\n\n\n\nexecute_tool<\/code> and invoke_connector<\/code> are the entire distance between “the agent did something weird overnight” and “the agent invoked this connector with these arguments at 03:14.” The pricing change is the sleeper of the three. Seat-based licensing on an agent platform is a direct tax on experimenting, and experimenting is the whole job right now.<\/p>\n\n\n\nQwen3.8-Max Is Not Open Source Yet, Whatever Monday’s Headlines Said<\/h2>\n\n\n\n
Sixty Percent of Enterprise Agents Were Handed Allow-All Access<\/h2>\n\n\n\n