Morning, humans.

Yesterday’s biggest AI number was not a benchmark score. It was $105 billion—the maximum guarantee Nvidia has agreed to provide for an Ohio data-center campus that OpenAI will lease for 20 years. If you were wondering when the AI race would stop looking like software and start looking like national infrastructure, this is your answer.

That deal leads AI news today, but the supporting cast is unusually revealing. Anthropic’s annual revenue run rate reportedly passed $65 billion. Microsoft faces questions over whether its installed chips match its public capacity claims. Voice startup Wispr raised $280 million. Meanwhile, OpenAI and Replit both pushed AI-assisted security closer to everyday engineering work. The theme is simple: AI’s winners increasingly need capital, power, distribution, and security—not merely a clever model and an energetic launch video.

AI news today: Nvidia underwrites OpenAI’s Ohio expansion

Nvidia will provide credit support of up to $105 billion for land, power, and shell construction at the PORTS-Pike Technology Campus in Pike County, Ohio. According to Nvidia’s announcement, SB Energy will build, own, and operate the site, while OpenAI becomes the customer under a 20-year lease. Nvidia will exclusively provide the AI compute infrastructure and invest another $1.5 billion in SB Energy.

The first phase targets 4.25 gigawatts of IT capacity, with an option covering another 3.75 gigawatts. OpenAI says the six-year buildout could create 35,000 construction jobs and 2,500 permanent roles. OpenAI and SB Energy have also committed a combined $80 million community fund.

The strategic point is larger than Ohio. Nvidia is helping finance the facilities that will buy Nvidia systems, while OpenAI secures compute without owning the campus. It is an elegant loop—assuming demand, electricity, construction, and financing all arrive on schedule. Infrastructure has a stubborn habit of ignoring demo-day timelines.

Anthropic’s reported run rate jumps past $65 billion

Anthropic’s annual revenue run rate exceeded $65 billion by the end of July, Reuters reported, citing a person familiar with financial updates shared with investors. That compares with $47 billion in May and roughly $9 billion at the end of 2025.

A run rate annualizes recent sales; it is not the same as audited full-year revenue. The figure is also reported rather than publicly confirmed by Anthropic. Even with those caveats, the acceleration helps explain why investors are entertaining extraordinary valuations ahead of a possible listing. Claude’s coding and enterprise adoption appear to be converting model capability into recurring spend—the part of the AI story that spreadsheets enjoy most.

Microsoft’s chip count raises a capacity question

A Guardian investigation says internal documents show Microsoft with about 2.2 million installed AI chips, despite roughly $280 billion in infrastructure spending since 2022. The publication argues that figure appears low relative to Microsoft’s statements about adding five gigawatts of data-center capacity.

Microsoft disputes the analysis, saying the estimates use incorrect assumptions, and notes that its facilities contain several chip types and generations. The company does not publish a detailed chip inventory, so outsiders cannot fully reconcile capacity, hardware, and utilization.

The useful takeaway is not that Microsoft has “run out” of chips. Buildings, grid connections, cooling, networking, and hardware must all become operational together. Announced gigawatts are not the same as usable compute. That gap is exactly why leadership teams need to separate AI ambition from AI readiness, a distinction explored in our AI readiness analysis.

Wispr raises $280 million to make typing optional

Voice-AI company Wispr announced a $280 million Series B at a $2 billion valuation, led by Menlo Ventures. The round takes its total capital raised to $361 million.

Wispr built its name around dictation, but the bet is broader: voice becomes an input layer across documents, messages, meetings, and software workflows. The funding matters because speech interfaces have spent decades being almost convenient. Better recognition, lower latency, and agentic follow-through could finally make them useful beyond setting kitchen timers.

Watch retention rather than download numbers. A voice product wins when people keep using it in noisy offices, with specialized vocabulary, and during real work—not when the demo understands one immaculate sentence.

OpenAI says the “defender’s window” is open

OpenAI president Greg Brockman published a security playbook arguing that organizations must begin automating more of their defensive work as models become better at finding and chaining vulnerabilities.

OpenAI says it is training models to produce more secure code, using AI to triage initial alerts, and connecting detections to bounded automated responses while reserving high-impact decisions for humans. Brockman recommends starting with approved, read-only assessments, reviewing internet-facing systems, and gradually moving toward code review and alert triage.

This is partly a vendor framing, and “superhumanly secure code” remains an aspiration rather than a demonstrated guarantee. The operational advice is sound: begin with constrained access, validate findings, and expand autonomy only after the system earns trust.

Replit adds an attacker’s-eye security test

Replit introduced black-box penetration testing for apps built on its platform. Unlike its existing scans, which inspect source code, the new test approaches an application through the network and browser without seeing its internals—closer to how an external attacker probes a live service.

That distinction matters as coding agents let smaller teams ship applications holding real customer data. Faster development compresses the time available for security review; embedding both white-box and black-box checks into the platform is a practical response. It does not replace expert testing for high-risk systems, but it can catch the unlocked doors that a code-only review misses.

For a broader view of how infrastructure and security pressures are converging, see yesterday’s AI and tech brief.

One thing to remember

The AI market is moving from model competition to systems competition. The companies that matter will secure electricity, financing, distribution, trusted revenue, and defensible software at the same time. Yesterday’s news showed how difficult—and expensive—that coordination is becoming.