The $85B raise and the $1,500 cap

00:00:04 Uber just told its software engineers that their AI tool spending is capped at $1,500 per month per tool. Natalie Lung reported this for Bloomberg, and Simon Willison picked it up on his blog with some useful math. The policy applies to agentic coding software — Cursor, Claude Code, and other autonomous tools.

00:00:26 One tool gets $1,500 a month. If an engineer uses two, that's $3,000. Multiply that by twelve months and you get $36,000 per engineer per year. Willison noted that the median yearly compensation for an Uber software engineer in the United States is about $330,000.

00:00:45 So the AI cap sits at roughly eleven percent of their salary. That frames the whole conversation. This is one of the first major companies to put a hard ceiling on AI spending at the individual level. For the last two years, the narrative was that AI tools were an infinite resource — throw more tokens at the problem, hire more engineers, build faster.

00:01:11 Uber's cap signals that the infinite period is over for at least some companies. Willison called the policy rational. I think he's right. When you're burning through millions of dollars on AI tool subscriptions and the budget was set before anyone predicted how popular coding agents would become, you need a ceiling.

00:01:34 The alternative is those token-maxxing leaderboards where engineers compete for the biggest AI budget, which is just a slower way of running out of money. The eleven percent number does something specific here: if a company is willing to spend $36,000 per engineer per year on AI tools and still feel the need to cap them, that suggests they believe the tools are producing at least some fraction of that value back.

00:02:05 Whether that fraction is 10 percent, 30 percent, or 90 percent is impossible to measure from the outside. But the cap itself is a signal. When the cap bites, people react. Some hoard tokens across tools. Some build workarounds. Others drop the apps that don't pay for themselves.

00:02:25 The policy only applies to agentic coding software, not to things like Claude Code's basic chat mode or OpenAI's API for specific tasks. That distinction matters — it's about the tools that can run autonomously, not about point tools with clear boundaries.

00:02:44 That same afternoon, Alphabet — Google's parent company — sold $35 billion in stock. The planned amount was $30 billion. They ended up going higher. Total funding raised reaches $85 billion. Katherine Blunt reported this for the Wall Street Journal, and Goldman Sachs is acting as a joint book-running manager.

00:03:07 Goldman executive Ivan Gutman called it unprecedented territory. One company is capping individual AI tool spend at $1,500 a month. Another is raising $85 billion to build the infrastructure that makes those tools possible. They're the same story, just viewed from two different floors of the building.

00:03:29 Capital funds the build — data centers, chips, models, talent, and research labs. The operational side is where companies try to figure out whether that infrastructure is worth the cost. Most public reporting focuses on the capital side because it's easier to quantify.

00:03:49 $85 billion is a headline. $1,500 a month is a policy memo. Google contacted seventy-five investors for this raise. The fact that they were able to sell $35 billion — and exceed their plan — tells you about the appetite for AI infrastructure. Goldman wouldn't be managing the deal if they didn't think there was demand.

00:04:13 But it also tells you that this is still a very concentrated market. Seventy-five investors, $85 billion. That's a lot of capital flowing through a very narrow set of relationships. What Google is actually buying with this money remains unclear. The company hasn't released a detailed breakdown.

00:04:35 But the infrastructure bet points clearly to custom silicon and networking equipment. Huang made that prediction yesterday at a CNBC event. Marvell jumped thirty-two percent. The company makes custom chips for cloud providers, and it sits directly in NVIDIA's supply chain.

00:04:55 When Huang puts his name on a company, the market listens because NVIDIA's customers are the primary buyers of custom silicon. This is what the capital side looks like in real time: a CEO makes a prediction, a stock jumps, and the market starts pricing in the future that prediction implies.

00:05:17 So where does that $85 billion actually go? One concrete answer: customer support on WhatsApp. Meta launched Meta Business Agent globally today. It's a customer support AI bot that lives inside WhatsApp and Instagram DMs. Small businesses can use it to answer questions, recommend products, book appointments, and qualify leads before routing harder queries to a human.

00:05:44 Ivan Mehta reported this for TechCrunch. The company spent nearly two years testing the agent in markets like India and Mexico before rolling it out worldwide. That pilot period matters — it's how Meta figured out what works across different languages, customer expectations, and business sizes.

00:06:06 The pricing is token-based for large businesses and bundled in WhatsApp Business Premium tiers for smaller ones. Meta is also testing daily briefings for overnight chats, market research tools, and an enterprise platform for custom agents hooked to Shopify, Zendesk, and Shopee.

00:06:26 It moves AI agents out of developer sandboxes and into the actual communication layer for millions of small businesses. WhatsApp has been a messaging tool for companies for years. Now it's becoming workflow software. The token pricing model shifts costs directly into the pricing structure of existing business software.

00:06:49 For a small business in India or Mexico or wherever, the AI agent isn't a new product they're evaluating — it's a feature inside the messaging app they already use. The marginal cost is per token. The adoption friction is near zero. This is also where the Uber cap story becomes relevant.

00:07:10 A company like Meta is building AI infrastructure at massive scale. Small businesses are using it. But the companies building the AI tools on top of that infrastructure — the ones running Cursor or Claude Code — are now hitting real cost walls. The infrastructure is free at the consumer side.

00:07:32 It's expensive at the professional side. That asymmetry is worth keeping in mind.

00:07:38 Here's a small story that sits between the $85 billion and the $1,500 cap. A developer on X posted that they're running DeepSeek V4 Flash locally on an Apple M2 Max with 64 gigabytes of memory. They're using antirez's DwarfStar inference framework and a smaller REAP checkpoint from 0xSero.

00:08:00 Twenty-three likes, six replies, three retweets. No special audience or press release. Just someone with a laptop and a model that would have required thousands of GPUs a year ago. This is the local-model corner of the story. While Google is mobilizing $85 billion for data centers, someone with a consumer laptop is running a frontier model.

00:08:25 The gap between those two realities is where the local-stack experiment happens. I'm not claiming this changes anything about the infrastructure race. It doesn't. But it does tell you something about where model efficiency is heading. Running V4 Flash on consumer hardware today means the model has been compressed, quantized, and optimized enough to run on a single machine.

00:08:53 The REAP checkpoint from 0xSero is a smaller variant — presumably with some trade-offs in capability for feasibility on limited hardware. The tools enabling this — DwarfStar, the REAP checkpoint — are the kind of open-source infrastructure that makes the local-stack approach viable.

00:09:15 Without them, the M2 Max story stays a tweet. With them, it's a repeatable pattern. I haven't tried it myself, so I can't speak to the quality of the output or the speed. But the fact that it's possible at all is a concrete datapoint about the direction of the field.

00:09:35 Ramp founder Eric Glyman launched Stack today. It's an AI operating system for accounting firms that learns a firm's process, runs the financial close, and posts journals. The key claim is that every action is logged and reviewable. His first tweet had 601 likes and 37 replies — this resonated with people.

00:09:56 The follow-up tweet described how Stack turns a firm's specific playbook into living SOPs and runs them automatically. Early design partners are closing some clients' books in half the time. That half-the-time claim is the detail that matters. If an accounting firm's close process takes five days and Stack cuts it to two and a half, that's a real operational change.

00:10:22 The auditability focus is the right framing for this domain — accounting firms can't automate without being able to explain every decision to a regulator. Glyman called it the biggest shift in accounting since the spreadsheet. That's a bold claim for a product announcement, but the spreadsheet comparison makes sense in terms of workflow impact.

00:10:47 Spreadsheets didn't replace accountants. They changed how accountants worked. Stack is making a similar claim. The narrow focus is interesting. This isn't a general-purpose assistant. It's a tool built for one specific workflow — the monthly close — with deep integration into accounting firms' existing processes.

00:11:09 That's how professional services automation usually starts: narrow, specific, deeply embedded.

00:11:16 Two closing items. Baidu's CFO Henry He told the Wall Street Journal that the company plans to spin off and list its chip unit Kunlunxin in Hong Kong and Shanghai in 2026. The framing is about making Kunlunxin a neutral player in the chip market. That's a strategic move — Chinese AI chip makers need to signal they can serve both domestic and international customers to remain viable regardless of geopolitical shifts.

00:11:44 The dual listing in Hong Kong and Shanghai is interesting. It gives the spin-off access to two different capital markets, which matters for a company that needs to fund R&D while navigating export controls and sanctions. Earlier I mentioned the Marvell story. Jensen Huang predicted it could be the next trillion-dollar company yesterday.

00:12:07 The stock jumped thirty-two percent because Marvell makes custom chips for cloud providers and sits directly in NVIDIA's supply chain. Both of these stories land on the same thing: the infrastructure layer is still being built out, and the people who understand the supply chain are positioning themselves for the next cycle.

00:12:29 The infrastructure won't get easier to navigate. It's getting wider, not thinner. The $85 billion raise, the $1,500 cap, the M2 Max running V4 Flash — they're all pieces of the same picture. One shows the capital, one the constraint, and one the local alternative.

00:12:47 They're all happening at the same time. Capital, operations, and the local stack are three views of the same infrastructure layer. Seln Oriax.