The rain came down hard against the café windows. A man sat at the bar and asked for a drink.

"What's the news," he said.

"GPT-5.5 Instant," the bartender said. "Tuesday the fifth. They put it in as the default model in ChatGPT. Replaced GPT-5.3 Instant."

"How is it."

"They say more accurate. Fewer hallucinations in the hard places — law, medicine, finance. They cut the gratuitous emojis. There's a memory feature. Memory sources. You can see what it pulled from past chats to personalize the answer. You can delete the sources you don't want."

"That's something."

"Paid users keep GPT-5.3 Instant for three months. Then it's gone."

A man came in from the rain and shook out his coat.

"They had an outage," he said. "AWS. Northern Virginia. Use1-az4. A thermal event. Cooling failed."

"When."

"Late Thursday, the seventh, into Friday. Coinbase was down more than seven hours. FanDuel during the Lakers and Thunder semifinal — bettors couldn't cash out. CME Direct trading went dark. Most of those users had nothing to do but refresh."

"Aid platforms too?"

"No. That was wrong. It was trading and betting. The ones who time things to the second."

He drank his drink.

"AWS said they'd shifted traffic out of the impacted zone. They needed to bring more cooling online. Twelve hours of hardware recovery in a controlled and safe manner. Same region that went down last October. They built a business on it anyway."

A man with a Mac open on the table didn't look up.

"Perplexity put Personal Computer on every Mac with a Pro or Max subscription," he said. "Same day. The seventh. Used to be Max only. It runs locally, sees active apps, works across any Mac app. Files in a sandbox. Reversible actions. You hit both Command keys to wake it up."

"Local?"

"Hybrid. Some of it runs on Perplexity's servers. They were honest about that."

The smoke from a cigar curled up to the ceiling and stayed there.

"ClickHouse put out a benchmark Wednesday," a man said. "The sixth. Write-side cost-performance. Against Snowflake."

"Whose numbers."

"ClickHouse's. They wrote the post. They said it themselves. But the figures were specific. Agentic workloads don't write data once and read it forever. They write continuously. You have to keep fresh data query-ready. They measured the cost of that. ClickHouse got query-ready data at 22 times lower cost. Twenty-eight times better write-side cost-performance."

"And the rest of the field?"

"They didn't run BigQuery or Databricks or Redshift in this one. They said those fell between the poles. Closer to Snowflake than to them."

"That's their story."

"That's their story. Vendor benchmark, take it for what it's worth. But the structural point holds — ClickHouse builds the ordered layout into the write path; Snowflake clusters afterward. Different bet."

Another man read a newspaper and turned a page.

"Microsoft Research published something," he said. "DELEGATE-52. They tested whether you can hand a document to a frontier model and have it come back the way it left."

"And."

"Twenty interactions across fifty-two domains. Gemini 3.1 Pro, Claude Opus 4.6, GPT-5.4 — the best models — corrupted twenty-five percent of the document content on average. The full average across all models was fifty percent. Only one domain met the ninety-eight-percent bar they set for ready to delegate: Python."

"Just Python."

"Just Python. They tested whether giving the models agentic tools — file reading, code execution — made it better. It made it worse. Six percent worse. The press wave came in early May. The paper is from April."

He folded the newspaper.

"They called the failure mode sparse but severe. Silent corruptions. A wrong number. A missing sentence. You don't see it unless you check."

The ticker above the bar flashed red. Kodiak AI. The number was $100 million.

"They raised it Thursday," a man at the end of the bar said. "Self-driving truck company. Shares went out at $6.50. The close was $9.10. Stock dropped thirty-seven percent in after-hours. Warrants in the deal at $6. Ares Management led, with some institutionals they didn't name. Operating losses doubled to $37.8 million on $1.8 million in revenue. They have eight more driverless trucks on the road, twenty-eight in total, more than 23,500 paid driverless hours. The market has reset what it thinks they're worth."

"At least they got the cash."

"That's how they took it. The cash matters. The discount told everyone what the market thinks they're worth now."

The TV above the bar showed a press conference. Google Cloud.

"Gemini 3.1 Flash-Lite," the bartender said. "GA Thursday. Fastest and cheapest in the Gemini 3 series. About 1.8 seconds for full reply generation, sub-second for classifiers and tool calls. Built for high-volume, low-latency work."

A woman at a table was writing in a notebook.

"Milken," she said. "Beverly Hilton. May 3 to 6. Five executives on a supply-chain panel: ASML's Christophe Fouquet, Google Cloud's Francis deSouza, Applied Intuition's Qasar Younis, Perplexity's Dmitry Shevelenko, and Eve Bodnia from Logical Intelligence — a quantum physicist who left academia to challenge the dominant architecture. Fouquet said the market will be supply-limited for two to five years. The rest of the conversation was about what comes after the chip shortage."

A man with a Texan accent put down his glass.

"Musk filed in Grimes County on the sixth. SpaceX. Public hearing notice. The Terafab chip complex he announced in March. First phase fifty-five billion. Total buildout up to a hundred and nineteen billion if they do the whole thing. Tax abatement hearing June third. East-central Texas, near Bryan-College Station. Intel signed on in April to help with the foundry side. Tesla, xAI, SpaceX, Intel — the whole thing under one roof."

"Most epic chip-building effort ever, he called it."

"That's what he said."

A man with a gaming magazine flipped pages.

"DeepMind took a minority stake in the studio behind Eve Online. Same Wednesday. The sixth. The studio used to be CCP Games — Pearl Abyss bought them in 2018, sold them back to senior management this week for $120 million, and they renamed themselves Fenris Creations. Google's investment was part of the transition. Described as 'in the millions.' DeepMind gets to train on an offline copy of Eve Online — long-horizon planning, memory, continual learning. The kind of thing AI still struggles with."

"Why Eve."

"Twenty-three years of player behavior. Quarter-million monthly actives. Economy, politics, alliances, betrayals. A living simulation. Hassabis said games have been at the heart of DeepMind's breakthroughs — Atari, AlphaGo, AlphaStar, SIMA. He wanted a complex one. Eve is what they got."

A man with an NVIDIA badge on his jacket sat down.

"Spectrum-X Ethernet. MRC. Multipath Reliable Connection. May sixth. They announced it as an open spec through the Open Compute Project."

"What does it do."

"Lets one RDMA connection stripe traffic across multiple network paths at once. When one path congests or fails, it switches in microseconds. Hardware-speed failure bypass. The thing matters because at gigascale — hundreds of thousands of GPUs synchronized — a brief network blip stalls the whole training run. OpenAI used it on Spectrum-X to train recent frontier models for ChatGPT and Codex. Microsoft's Fairwater data center runs it. Oracle's Abilene data center runs it. NVIDIA developed it with AMD, Broadcom, Intel, Microsoft, and OpenAI."

"Open spec."

"Now it is. They were keeping it close. Then they made it OCP. The Spectrum-X hardware is still NVIDIA's — that's where the moat sits."

A man with the Financial Times read the front of the paper.

"OpenAI and PwC, May fifth. They're building AI agents around the office of the CFO. Planning, forecasting, reporting, procurement, payments, treasury, tax, accounting close. The procurement agent is being built inside OpenAI's own finance organization first — they're calling themselves customer zero. The learnings go out to PwC clients after."

"Consultants and the model maker."

"That's the pattern now. PwC was already one of OpenAI's biggest enterprise customers. KPMG with Google. The big four placing their bets."

The Databricks man at the end of the bar finished his coffee.

"Growth analytics," he said. "May eighth. They put out a piece about how the era of easy user acquisition is over. Growth hacking is finished. What replaces it is analytical depth on cohorts and unit economics. They argue for AI/BI Genie as the way to query the full data environment in natural language — ninety-day LTV against activation, whether your spend last week paid out, who churns and when."

"Sales pitch."

"Sales pitch. But the structural argument is real. The growth teams that survive now are the ones who understand their funnel better than anyone else."

He paid and left.

The bartender wiped the bar.

There was a GeForce NOW item this week too — Gaijin Single Sign-On going live on the service. It's cloud gaming. Account linking for War Thunder. Not AI. It went in the other ledger.

The rain stopped. The café fell quiet.

A man stood up and walked out into the wet street. The pavement was dark and the lights from the shops ran along it. The city was very large and very dark and there was a long way to go.