AI & Finance Weekly

Issue 26 · Monday, Jun 29 2026

Get fluent in AI — and in what it means for your numbers.

A five-minute Monday brief that teaches the AI behind each story, then translates it for the people who actually close the books. Every number links to its source.

InsideModelsEnterpriseMoney & Rules

3 things to know this week

Google had its worst market day in over a year after two AI stars left for OpenAI and Anthropic.

Oracle logged its worst week since 2001 as investors recoiled at the bill for its AI buildout.

A free Chinese model, GLM-5.2, drew enterprises just as Washington gated US frontier models.

$60B

▲ all-stock

SpaceX's price for Cursor-maker Anysphere

−$24B

▼ free cash flow

Oracle's FCF last fiscal year, amid AI capex

$130B

▲ debt load

Oracle's debt at the end of May

⅕

▼ the cost

GLM-5.2 vs Opus 4.8, near parity on a key test

Lead story · Alphabet

Google had its worst market day in over a year — after two of its AI stars walked out

Per CNBC, Alphabet had its worst trading day in over a year on June 22, falling roughly 5%, after two of its most prominent AI researchers left for rivals: Noam Shazeer — a Gemini co-lead and co-author of the foundational "Attention Is All You Need" paper — went to OpenAI on June 18, and John Jumper, the Nobel laureate who led DeepMind's AlphaFold, announced a move to Anthropic on June 19. Investors weren't reacting to a product; they were repricing whether Google can hold onto the small number of people who build a frontier model.

The concept · key-person risk

At the frontier, a model's edge rides on a handful of researchers who know how to turn compute and data into a working system. That's a familiar risk in finance: when value is concentrated in a few people, losing them can re-rate the whole asset overnight.

Read the report on CNBC →

Modelslabs & frontier

Models · open weights

Zhipu (Z.ai) · GLM-5.2 · Jun 2026

A free Chinese model is now within a point of the best US model — at about a fifth of the cost

Per CNBC, Zhipu's open-source GLM-5.2 lands within roughly one percentage point of Anthropic's Opus 4.8 on a closely watched agentic benchmark, at about a fifth of the price — and it's free to download, fine-tune, and run on your own servers. Demand surged just as US frontier access started to look shaky: Anthropic's Fable and Mythos models were pulled under a government order, and OpenAI said it's limiting GPT-5.6 at Washington's request. One enterprise AI leader told CNBC it's the first open model that's genuinely competitive with the closed frontier.

The concept · open weights

A closed model lives behind a vendor's API, which can be throttled, repriced, or switched off. An open-weight model ships the actual file — download it once and no government or vendor can revoke your copy.

For your desk

This is last week's Fable 5 lesson with a price tag attached: "intelligence per dollar" and "can't be revoked" are becoming real procurement criteria. Worth knowing which of your AI workflows could fall back to a self-hosted model if a hosted one went dark or jumped in price.

Read the report on CNBC →

Enterpriseplatforms & the market

Enterprise · M&A

SpaceX / Anysphere · acquisition · Jun 16

SpaceX is buying Cursor for $60B — the clearest sign yet that coding agents are real AI revenue

Per CNBC, SpaceX agreed to acquire Anysphere — maker of the AI coding agent Cursor — for $60 billion in an all-stock deal, days after its record IPO. It's a bet on AI-assisted coding, one of the first places AI has turned into real enterprise revenue. The twist: Cursor's share of that market had already slipped from about 41% a year ago to roughly 26% by May, with Anthropic's Claude Code taking close to half the category, according to Ramp spending data cited by CNBC.

The concept · vertical integration

Buying a layer of your own supply chain — here, a customer-facing app that sits on top of your models and compute — to capture more of the value, and the user data, instead of letting a partner keep it.

For your desk

Consolidation is coming to the AI tools you already pay for. When a model lab or a hyperscaler buys your coding or analytics vendor, pricing, data terms, and the roadmap can all shift — worth re-reading the renewal and change-of-control clauses before they do.

Read the report on CNBC →

Money & Rulescapital & regulation

Money · markets

Oracle · markets · Jun 26

Oracle had its worst week since the dot-com bust — because investors did the AI math

Per CNBC, Oracle logged its worst week since 2001 as investors focused on what its AI buildout costs: capital spending jumped 162% to about $56 billion last fiscal year, free cash flow ran to roughly negative $24 billion, and debt sat near $130 billion at the end of May — with another $40 billion in debt and equity planned. The stock is down about 24% in 2026. The demand for AI compute is real; the question Wall Street is suddenly asking out loud is whether spending on this scale pays back.

The concept · capex vs. free cash flow

Capital spending builds the data centers now; the revenue — and the depreciation — arrive over years. Negative free cash flow means a company is paying out far more cash than it takes in: fine if the payoff lands, painful if it slips.

For your desk

This is the exact tension behind every AI business case you'll review — the spend lands now, the benefit later. When the AI line shows up in a capex plan, "what's the payback period, and what happens if it slips a year?" is the question to have ready.

Read the report on CNBC →

AI, demystified

What it actually takes to build a frontier model

This week's three big stories — a talent exodus, a capex scare, and a cheap open model — are all really about the same three scarce inputs. Here's the whole machine on one page.

Compute

The data-center power to train and run the model. It's the biggest cash cost — and the reason Oracle's bill rattled investors.

The factory floor.

Data

What the model learns from. More — and cleaner — data makes it sharper, which is how a cheap open model can suddenly catch up.

The raw material.

Talent

The few researchers who know how to turn compute and data into a working frontier model. Lose a couple and the edge wobbles.

The engineers who run it.

Every headline this week is one of these three getting scarcer or pricier: Google lost talent, Oracle is straining on compute, and GLM-5.2 used data efficiency to undercut everyone.

Hype check Exaggerated "Google has lost the AI race."

The line going around after this week's exits. Two marquee researchers leaving in a week is a real blow — and the market said so, with Alphabet's worst day in over a year. But "lost the race" doesn't survive the details: Gemini still powers Search, Workspace and Android for billions of people today, DeepMind still employs thousands of researchers, and Alphabet's profitable core can fund AI spending for years without raising outside capital.

The honest read is narrower — a retention problem and a momentum question, not a knockout. Losing the people who helped invent the Transformer and AlphaFold is exactly the kind of signal investors watch. It changes the odds; it doesn't settle the outcome.

Reality-check source: CNBC ↗

Do this week

Map your AI single points of failure.

List the workflows that depend on one model or one vendor.
For each, note whether there's a fallback you could switch to this quarter.
Flag any that would simply stop if access were pulled or repriced — that's your risk list.

~20 min · the Fable 5 / GLM-5.2 lesson, made concrete

Prompt of the week

Make AI surface your vendor dependencies — and its own uncertainty:

# paste a vendor contract or a 10-K risk-factors section
List every dependency on a single AI model or provider, the switching cost if it changed, and any termination or price-change clause. Flag anything you're unsure about instead of guessing.

Teaches the habit: ask AI to expose dependencies — and what it isn't sure of.

Number of the week

How much Oracle's capital spending jumped last fiscal year — to about $56 billion — to build AI data centers. The demand is real; the worry now showing up in the share price is whether spending like this pays back, and how soon.

Source: CNBC ↗

Off the recordfile under: the talent war

Off the record

Why a Nobel laureate and a Transformer co-author walked out of the most storied lab in AI

DeepMind used to brag that nobody left. This month two of its biggest names did, in 48 hours — Shazeer to OpenAI, Jumper to Anthropic. Per Fortune, the simplest explanation, money, doesn't quite fit: both were already wealthy. The likelier pull is the chance to ship faster, with fewer layers of approval, at labs that do one thing instead of running search, ads, Android and a cloud all at once.

Real talk: it's the same reason your best people leave — rarely the pay, usually whether they can actually get things done. The team that strips out friction is the one that keeps its talent.

Personality of the week

No. 26 · the Nobel laureate

Why this person, this week

John Jumper

Joining Anthropic · ex-Google DeepMind

The most decorated scientist in AI just bet on where the frontier is headed.

Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold, the system that predicts protein structures — arguably AI's biggest scientific breakthrough to date. On June 19 he announced he's leaving DeepMind after nearly nine years to join Anthropic, in what's been called the highest-profile AI talent move of the year. His focus — AI for science — is a tell about where serious researchers think the next breakthroughs, and the next markets, will come from.

Read the story →

Until next Monday — keep a name next to every number.

Every figure links to its source — numbers are as reported by the cited outlet; verify before you cite them onward.