The number dominating the AI feed this morning is $250 billion.
According to a Wall Street Journal report this weekend — confirmed by Reuters and Bloomberg — Nvidia is in talks to provide a $250 billion financing backstop for OpenAI's planned data center campus in Piketon, Ohio. The campus itself could cost more than $500 billion to build. Separately, Nvidia is reportedly discussing financing for OpenAI's chip purchases worth up to another $350 billion. Total exposure being discussed: somewhere in the neighborhood of $600 billion. On a former uranium enrichment plant in southern Ohio.
I want to sit with that number for a moment before we talk about what it means for the law firms, financial advisory practices, consulting firms, and healthcare practices that actually make up The Promptory's audience.
Because the gap between what the AI industry is spending and what your firm needs to spend to get real, measurable results from AI is the most underreported story in technology right now.
The Disconnect Nobody Is Naming
The AI industry is spending $700 billion on infrastructure this year. Most professional services firms still can't answer whether their $300/month in AI subscriptions is doing anything.
Both things are true at the same time. The scale of AI infrastructure investment is real and staggering. The gap between that investment and measurable business outcomes at the firm level is equally real. This week's newsletter is about closing the second gap — because that's the one you can actually do something about.
What's Actually Happening This Week — Three Stories, One Honest Translation
WSJ, July 26 · confirmed by Reuters and Bloomberg · not yet verified by Nvidia or OpenAI.
The reported deal structure: Nvidia would guarantee financing tied to the data center lease and construction debt for a planned 10-gigawatt campus in southern Ohio, developed by SoftBank's energy subsidiary. The reason Nvidia's guarantee is needed at all is telling — OpenAI, despite being valued at over $850 billion and projecting around $25 billion in 2026 revenue, lacks an investment-grade credit rating. Conventional lenders declined to back the project directly. Nvidia stepped in because guaranteeing OpenAI's infrastructure also guarantees demand for Nvidia chips for years.
Michael Burry reacted on X: "Around and around we go." He's been building a short position in Nvidia. Analysts are calling it circular financing — Nvidia invests in OpenAI, OpenAI buys Nvidia chips with the money. The circular concern is real and worth noting. This doesn't mean AI infrastructure spending isn't justified — it means the financial structure behind some of it is less solid than the press releases suggest.
What it means for your firm: The infrastructure being built in Piketon is what makes the AI tools you use cheaper and more capable over time. Inference costs have already fallen dramatically — the tools available to a 10-person professional services firm today would have cost 100x more two years ago. That trend continues regardless of how the financing structures shake out. For your firm, cheaper is already here.
Anthropic · July 24, 2026 · $5/$25 per million tokens · same price as Opus 4.8 · powers Jordan.
Anthropic launched Claude Opus 5 Thursday — near-frontier intelligence at the same price as the model it replaced. The benchmark headline: Opus 5 scored 43.3% on FrontierBench v0.1, compared to GPT-5.6 Sol's 37.5%, putting Anthropic at the top of that leaderboard. On Zapier's internal AutomationBench, Opus 5 completed a full churn-prevention workflow start to finish — Wade Foster, Zapier's CEO, said prior models didn't pass and Opus 5 hit 100%.
The more useful framing came from Anthropic product lead Dianne Penn, who told Reuters the rough rule is Opus 5 for "complex but routine work," and the pricier Fable 5 for "days-long, very autonomous projects." In plain terms: Opus 5 is built for the work that fills a professional services week — research synthesis, reporting, multi-step client deliverables, document review.
One honest note: an independent review found Opus 5's hallucination rate is about 14 percentage points higher than Opus 4.8's. Customer feedback highlighted that it "behaves more like a careful professional than a text generator" — but for any AI-assisted work product leaving your firm, human review remains essential. The model improved; the review requirement didn't disappear.
What it means for your firm: Jordan runs on Claude. When Anthropic's flagship model gets meaningfully better at the same price, Jordan's diagnostic quality improves with it — the bottleneck analysis, the tool recommendations, the implementation scoping all get sharper.
Moonshot AI · July 27, 2026 · 2.8 trillion parameters · free to download · 1.4TB full weights.
Moonshot AI's Kimi K3 open weights went live at midnight UTC — making the 2.8-trillion-parameter model free to download and self-host. Self-hosting at that scale rules it out for most small firms. But the signal matters regardless of whether you're downloading it.
What it means for your firm: Open-weight releases at this scale apply downward pricing pressure on every closed model API. When a model that topped a coding leaderboard becomes free to download, it changes the competitive calculus for OpenAI, Anthropic, and Google — they respond by making their own models cheaper or more capable, or both. The trend line for what you can access for $50/month is moving in one direction.
What All of This Actually Means for Professional Services
Strip out the infrastructure awe and the benchmark leaderboards, and here's what this week's news cycle is actually telling professional services firms.
The capability is no longer the constraint. Claude Opus 5 at the same price as its predecessor. Kimi K3 free to download. DeepSeek at $0.44 per million output tokens. The model performance available to your firm right now — at a price that rounds to zero against your other operating costs — is extraordinary. The firms not seeing results from AI aren't being held back by model quality. They never were.
The implementation gap is widening, not closing. The more capable the models get, the larger the gap between what AI can do and what any given firm has built the systems to capture. Every week a firm runs subscriptions without defined workflows, without measured outcomes, without an implementation layer connecting the tool to the work, is a week the gap grows.
The $250 billion and the $300/month are not the same conversation. The Nvidia-OpenAI financing story matters for understanding where the industry is heading. It doesn't tell you what to do this week in your firm. The firms getting real, measurable results aren't doing it with frontier research infrastructure — they're doing it with a CRM that actually runs, a follow-up sequence that actually fires, a policy that actually governs what their team does, and a clear metric that tells them whether it's working.
The firms asking the right question are pulling ahead. The right question isn't "which model should we use?" It's "what specific problem are we solving, what does success look like in 30 days, and who is accountable for making sure it happens?" That question — not model selection — is what separates the 7% of organizations with established AI ROI from the 93% still waiting to see something on the bottom line.
The Four Pain Points We Hear Every Week
This is the most common sentence Jordan hears. It's not about the tools — it's the absence of a success metric defined before the tools were deployed. The fix isn't a new tool. It's a 30-minute conversation that answers: what specific outcome were we trying to move, what's the number, and are we tracking it? No metric was defined before launch — the solution is a retrospective definition, then measurement starting now.
This is a workflow integration failure, not a people failure. When a new tool is added alongside the existing process — rather than inside it — reverting is the path of least resistance. MIT NANDA research was direct: generic AI tools "do not adapt to existing workflows." The tool needs to replace a step in the existing process, not sit next to it as an optional extra. Implementation stopped at configuration and never reached workflow redesign — the fix is removing the old path, not adding a new one alongside it.
This one is accelerating. Enterprise procurement teams are adding AI governance requirements to vendor questionnaires. A firm without a clear, documented AI policy is increasingly at a disadvantage in enterprise client relationships — not just a compliance risk. A one-page policy, written this week, closes most of the gap. The template ran in Issue #038 of this newsletter.
This is the curation problem The Promptory was built to solve. There are now more than 50,000 AI tools on the market, with new launches and benchmarks every week. Model selection — GPT or Claude or Gemini, Opus 5 or Fable 5 — isn't the question that determines results. The question is which tool, configured correctly, connected to which specific workflow, measured against which outcome. Tool selection is happening before problem definition — the vault gives you curated, vetted tools, and Jordan gives you the problem definition that tells you which ones apply.
Ask Jordan
I just read about Claude Opus 5, the Nvidia deal, Kimi K3 — and I feel like I'm watching the AI world spend incomprehensible amounts of money while I'm sitting here not sure if my $200/month in ChatGPT and Copilot subscriptions is doing anything useful. Where does a firm like mine actually start?
Exactly the right feeling to name — and a very common one right now. The $250 billion and your $200/month are not the same conversation. You start here: tell me the one task in your firm that happens most often, takes the most time, and produces the least value for what it costs your team. Not a category — one actual task. That answer tells me everything I need to know about where to start. The model question comes last, not first.
The AI industry is spending $700 billion on infrastructure this year. You don't need any of it to get results. You need a defined problem, a measurable outcome, and a system built to move them.
Claude Opus 5 is genuinely better than anything available six months ago, at the same price. Kimi K3 is free. Inference costs are falling. The capability available to a professional services firm right now is extraordinary and getting cheaper. The constraint has never been the technology. It's always been the problem definition, the workflow integration, and the metric that tells you whether it's working.
The firms that close that gap this month will be a full quarter ahead by the time the Ohio data center comes online. The infrastructure being built in Piketon will make their already-working systems faster and cheaper. It won't do anything for the firms that still haven't defined the problem.