Uber's full-year AI budget was gone by April

This was the week AI stopped being priced like magic and started being priced like electricity — enormous at the power plant, pocket change at the plug.

→ Nvidia's $250B co-sign — the chip seller moves to guarantee OpenAI's Ohio mega-lease, and the cost of insuring Nvidia's own debt hits a record
→ The agent bill — Uber burned its full-year AI budget in four months; a research lab just published the formula for when agents lose you money
→ The paycheck split — AI skills now pay 62% more, while US tech and finance shed 28,000 jobs a month
→ Pocket-change tools — Cursor at $7/month in India, Grok free inside Google Docs and Sheets, DeepSeek under $1

Surveyors are staking a field in Piketon, Ohio — a decommissioned uranium-enrichment site — for a campus that would draw 10 gigawatts of power. On Monday, the price traders pay to insure Nvidia's debt hit a record — the biggest one-day jump since the contract began trading last November.

Why? Lenders won't fund this on OpenAI's numbers alone: an $852 billion valuation, still unprofitable, a ~$14 billion loss projected this year. So Nvidia is in early talks to guarantee the lease debt — if OpenAI can't pay, Nvidia pays.

Also on the table: up to $350 billion more to finance OpenAI's purchases of Nvidia's own chips — the seller, funding the buyer's shopping.

What it means for you:

→ Two-minute check, no Bloomberg terminal needed: Google your fund's name plus "factsheet" and open the top-10 holdings. If one chip company is above ~10% of your fund, you're making a concentrated AI bet — fine, but decide it on purpose. → It's early talks — a collapse costs nothing, but a signed deal makes Nvidia's balance sheet the safety net for the entire closed-AI economy. Watch the guarantee, not the $500B headline. → This week, when you hear "circular financing" — in a meeting, a headline, a group chat — you can translate it: the seller co-signing the buyer's loans.

Does it actually pay for itself? The same question lands on a smaller desk — yours — when the first AI agent invoice arrives.

In April, Uber's finance team opened the AI coding-tools line: the entire 2026 budget, gone — twelve months of spend in four. Claude Code, Anthropic's coding agent, had gone from 32% of engineers to 84% in one month—the fix: a $1,500-a-month cap per employee.

Why it was a surprise matters to everyone:

→ Agents are metered per token — the word-chunks AI bills by, like kilowatt-hours — so the invoice arrives after the decision, unlike a subscription.
→ Agents burn 10-100x more tokens than a chatbot: every step re-sends the whole conversation so far.
→ The expenditure horizon, named last week by research lab METR: the per-task budget below which an agent beats a human — spend more than that on one task and the human wins. For today's best models: about $2,300-$3,300 per task.

Companies are running this same math right now — except the line item isn't software. It's salaries.

A software company still forecasting 19-20% growth this year: the all-hands doc announces some 620 layoffs — 20% of the company. The same day, Amazon cuts jobs in its own AGI division — the team building its most advanced AI.

The other letter going out this year is an offer letter. PwC scanned a billion-plus job ads; the US government counted the losses:

What it means for you:

→ "Profitable and growing" no longer means safe; visibly AI-skilled increasingly does — a workflow you can demo, not "AI" on a resume. → Roles are merging, not just vanishing: 43.5% of occupation-specific work messages to ChatGPT involved another occupation's tasks, per OpenAI's own study. The adjacent skill is the hedge. → Paste this: "Here are my five most time-consuming weekly tasks: [list]. Pick the one where AI saves the most time and give me a 30-minute first version with tools I already have." Two minutes to get the pick, thirty to build — talk about it after it works.

The premium goes to people who visibly do things with AI. The good news: those tools just collapsed in price.

At a Bengaluru chai stall, a developer does the math: chai, 20 rupees; Cursor — the AI code editor whose third-largest market is India — now about $7 (₹649) a month. At her desk, Grok sits in her Google Sheet. Free.

→ Cursor Start (India-only, launched Monday, UPI accepted): Composer 2.5 plus Grok 4.5 — capable current-generation models, enough for real AI-assisted coding. OpenAI's and Anthropic's top models stay on the $20 plan: $7 buys the workflow, not the flagship.
→ Grok for Google Workspace: $0 — answers questions from your Sheets with cells cited, writes formulas, builds Slides from outlines.
→ DeepSeek V4: under $1 per million tokens of output — roughly 750,000 words — after a ~75% cut. Surge pricing for Beijing business hours is announced for the full release. The driver: a Chinese open-model price war, and Cursor chasing India's 27 million GitHub developers.

What it means for you:

→ Install the Grok add-on from the Google Workspace Marketplace — two minutes, $0, zero new apps. → Write code in India? Take the $7 plan; upgrade when you need the top models. → Buying AI through an API? Get a DeepSeek quote before your next renewal — even if you never switch, it's leverage.

The week Nvidia moved to co-sign $250 billion at the grid, the plug price fell to $7. Know the cost, know the worth — the math favors you.