TOGETHER WITH TODAY'S PARTNER
Good morning, {{first_name|there}}. Bond traders just put a number on the promises holding up the AI buildout: $70 billion that appears on nobody's balance sheet.
Read time: 3 minutes. Same time, every weekday — rate today's issue at the bottom.
🚀 The Big Story: The AI boom's $70B of shadow credit
Bloomberg reports bond investors are now tracking roughly $70 billion in residual value guarantees — chipmakers quietly promising to eat the losses if AI borrowers default or if the hardware depreciates faster than the loan.
Meta wrote the template: $27B (Beignet) and $13B (Sopaipilla) data center packages financed with these guarantees, keeping the liability off the balance sheet.
Broadcom backstopped $35B of chip financing for Anthropic's "Project Big Sky" — a structure analysts say could stack up to $370B in senior debt by mid-2029.
Nvidia is writing puts: it may provide residual-value support on up to 25% of its $500B financing partnership with BlackRock, Goldman, Apollo, Blackstone, Brookfield, and KKR.
The design flaw: these guarantees are pro-cyclical. They trigger exactly when customers are defaulting and used GPU prices are collapsing — the worst possible moment for the guarantor.
Jason's take: Your token prices are a financing product, not a technology price. When a chipmaker guarantees the loan that bought the GPU that serves your API call, the number on your invoice reflects somebody's balance sheet appetite — not the cost of inference. DeepSeek raising API prices up to 1,100% today is what the repricing looks like when appetite thins. Build like your cheapest model doubles in Q4, because one of them will.
⚡ Quick Hits
Anthropic booked $11.5B in Q2. Up from $787M a year ago — 14x — with its first positive adjusted operating income. Backers are now floating a $2T valuation.
Qwen passed 3 billion downloads. Alibaba's open models beat Google (418M) and Meta (227M) combined, with 300,000+ derivatives. Hugging Face calls Qwen "part of the default workflow."
DeepSeek V4-Pro repriced today. API costs jump up to 1,100% as of August 16 — the cheap-token era's clearest obituary yet.
Gemini 3.7 Flash closed the coding gap. FrontierCode 1.1 went 34.4% → 43.6%, DeepSWE 49% → 65.3%, at $0.75 per million input tokens through December 31.
Nvidia's 13F is a venture fund. $29.99B in Intel and $20.98B in SpaceX are 80% of its $63.4B equity book. The supplier is now the shareholder.
Anthropic shelved "Model 2." An internal system described as materially more capable than Mythos 5 stayed in the box as the lab raised its catastrophic-misalignment rating from "very low" to "low."
📡 Trending on X
"Adjusted operating income" is the fight of the weekend. Ed Zitron's "Anthropic's Profitability Swindle" is everywhere, and the replies are a masterclass in what that adjective is carrying.
American open source is having a bad news cycle. The Qwen numbers turned into a referendum on Llama's retreat, with "the default open model is Chinese now" as the quote-tweet of the day.
DeepSeek's price hike broke the goodwill. Builders who ported production traffic over on price are posting migration threads and cost screenshots, and the mood is betrayal.
"Nvidia is writing puts on its own customers." The circular-financing crowd got its best evidence yet, and the bulls are arguing that a guarantee you never pay is just good sales.
📺 Trending on YouTube
🛠 The Workflow: The 15-minute price-shock audit
DeepSeek just repriced overnight. Here's how to find out what that would cost you before it happens to your provider:
Pull last month's bill from every model provider and write down input and output tokens per endpoint, not just dollars.
Multiply by 3. If that number kills the unit economics of a feature, that feature is subsidy-dependent — mark it.
Route the marked features to two providers. One frontier, one open-weight you could self-host. Qwen 3.8 and Kimi K3 both have Apache-style licenses.
Write the eval before you need it. Twenty real prompts with graded outputs, so a provider swap is a Tuesday, not a project.
Cache and trim. Prompt caching plus cutting your system prompt usually recovers 30–50% before you change a single model.
Reply with the word "AUDIT" and I'll send you the spreadsheet I use, with the eval template built in.
🧰 Trending Tools
Kilo Code — reviews and refactors like a senior dev, not an autocomplete. For solo builders shipping without a team to review the PR.
Google Stitch 2.0 — text to production UI for dashboards and landing pages. For operators who lose weeks waiting on design.
Fathom 3.0 — records and summarizes calls with action items attached. For anyone whose follow-up quality decides whether the deal closes.
AdAnt AI — generates and iterates social ad concepts at volume. For marketers who need 30 variants by Thursday.
Hey Noah — an AI chief of staff for calendar, inbox, and relationship follow-up. For consultants whose pipeline dies in the gaps.
📣 Put your brand here. The AI Innovator reaches AI-first operators, creators, and marketers every weekday. Primary sponsorships are now booking.
That's a wrap
Tomorrow: what Anthropic's first profitable quarter actually costs you at renewal.
Enjoying the new format? Share your link — every referral keeps this free:
How was today’s email?
You're reading the 3-minute AI Innovator — same time, every weekday. Hit reply and tell me what you want more of. I read every one.
