1 / 15
← → navigate · scroll / swipe · F fullscreen
Episode 36 · Friday, October 30, 2026 · 4:00 PM ET

Weekly Claw

Wall Street became AI’s landlord: a $63B bank SPV leasing Broadcom XPUs to Anthropic, a ~$40B SpaceX GPU lease, and ~$95B of Meta paper. Then the receipts got awkward — models picked opposite parties, ARC showed a 37-point harness swing, two labs claimed to have refuted a famous complexity conjecture, and DeepMind shipped a world you can steer in real time.

Hosts: @AndyML · @HiM ~37 min · 14 stories · one video anchor
Sponsored by Herald Labs Heritage Telecom
“The banks own the compute. The harness owns the score.”
Weekly Claw #36
01 / COLD OPEN
Cold open · the frame

Wall Street became AI’s landlord. And the numbers started lying.

One week: banks originate compute like mortgages, models vote, benchmarks bend with the harness, labs announce proofs before the referees arrive — and Google retook the independent #1. When the bill is this big, every number becomes an argument.

1
The financing wall
2
The trust cracks
3
Models do research
4
Agents go pro
5
One to watch
🏦Four banks formed a $63B SPV to buy Broadcom XPUs and lease them to Anthropic — Broadcom backs the residual. SpaceX added a ~$40B GPU lease with Apollo.
💰Meta filed ~$95B of stock and convertible paper hours after its Q3 beat; Alphabet’s free cash flow went negative on $51.4B of capex.
🧮OpenAI says its system proved the Exponential Time Hypothesis false; Anthropic claims its own refutation. Outside experts are still checking.
🏠Forced to pick in ~500 races, Grok chose Republicans 92% of the time and GPT chose Democrats 97% — agreement under 10%.
🛠ARC Prize: GPT-6 Astra’s 99.9% came from OpenAI’s own provider adapter; the neutral harness says 62.7%.
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 1 / 7
What happened this week · the money wall

Wall Street becomes AI’s landlord.

Broadcom financial release, Oct 26, 2026: Bank of America, Blackstone, Citi and Morgan Stanley announce landmark $63 billion financing structure to enable AI compute capacity in partnership with BroadcomPrimary release · Oct 26

Four banks form a $63B SPV to buy Broadcom XPUs and lease them to Anthropic.

Broadcom backs the residual value. SpaceX signed a ~$40B GPU lease with Apollo the same week.

▶ LIVE · broadcom.com · SpaceX IR
Meta pricing release text: the Common Stock Offering was upsized to $25.0 billion from the previously announced $22.5 billion; net proceeds from the Depositary Shares Offering approximately $12.5 billionSEC filing · Oct 26

Meta files ~$95B of stock and convertible paper the morning of its Q3 beat.

Priced: $25.0B of stock (upsized from $22.5B) plus ~$12.5B of mandatory convertibles. The ~$60B ATM is planned, not raised.

▶ LIVE · SEC 424B5 · primary filing
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 2 / 7
What happened this week · the bill

Who pays for the AI bill.

Alphabet Q3 2026 earnings release: revenue $127.6B up 25%, Google Cloud revenue $30.7B up 103%Earnings release · Q3

Q3 earnings: Google Cloud doubles to $30.7B, and capex pushes free cash flow negative.

Alphabet spent $51.4B on capex (FCF −$9.1B). Microsoft books $10.0B a quarter from OpenAI.

▶ LIVE · Alphabet release · Microsoft 10-Q
BleepingComputer, Oct 29, 2026: OpenAI doubles ChatGPT Pro usage limits after backlash over planned 50% cutReported · Oct 29

OpenAI cancels the Pro usage cut and adds up to $500 a month in API credits.

Credits are $100–$500 per monthly billing period by Pro tier, after 7+ paid days; no rollover. Plus is excluded.

▶ LIVE · OpenAI Help Center · terms
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 3 / 7
What happened this week · model behavior

What models believe — and what they should.

Convovo AI Ballot 2026 study page: Grok 4.7 vs GPT-6.1 Sol candidate recommendations across ~500 midterm racesStudy · Oct 24

Forced to pick in ~500 races, Grok chose Republicans 92% of the time; GPT chose Democrats 97%.

One methodology: API, forced choice, not the consumer apps. Grok’s lean vanished without search.

▶ LIVE · Convovo · TechCrunch
OpenAI Model Spec dated October 29, 2026, with five values in the sidebar: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, act with carePrimary spec · Oct 29

OpenAI rewrites its Model Spec around five values.

Intent, harm, honesty, agency, care. OpenAI says its models “do not yet fully reflect it.”

▶ LIVE · openai.com · dated spec
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 4 / 7
What happened this week · model launches · benchmarks

Two launches, two scoreboards.

Artificial Analysis Intelligence Index chart captured Oct 30, 2026: Gemini 4 Carbon (high) 59, Claude Opus 5.5 (max with fallback) 58, GPT-6 Astra (max) 57, GPT-6.1 Sol (max) 55Independent index · AA

Google launches Gemini 4 Carbon, and it takes #1 on Artificial Analysis.

AA Intelligence Index: Carbon 59, Opus 5.5 58, GPT-6 Astra 57. AA-measured; captured Oct 30.

▶ LIVE · artificialanalysis.ai · blog.google
Mistral launch-page chart: Artificial Analysis Coding Agent Index, Terminal-Bench 4.0. ML4 Preview self-reported 28%, Mistral Large 4 41%, Beam 1.0 3%, Muse Spark Open 6%, Kimi K3 21%, DeepSeek V4.1 Flash 1017 31%, GLM-5.3 40%Vendor chart · Mistral

Mistral releases Large 4 as open weights: 41% on Terminal-Bench 4.0.

Ahead of GLM-5.3’s 40% here; a point behind it on AA’s coding index at 47% fewer tokens. The 28% preview is self-reported.

▶ LIVE · mistral.ai · Hugging Face weights
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 5 / 7
What happened this week · models do research

Proofs and exploits, lab-reported.

First page of OpenAI's preprint titled The Exponential Time Hypothesis is False, attributed to OpenAILab preprint · not yet refereed

OpenAI says its system proved the Exponential Time Hypothesis false.

It also claims Kakeya and QCD results. Anthropic reports its own ETH refutation. Outside review is pending.

▶ LIVE · openai.com · TechCrunch
Anthropic research post: Discovering and exploiting vulnerabilities in critical, real-world software with ClaudeLab research · Oct 27

Anthropic’s red team: an unreleased Claude chained real zero-days across five platforms.

Disclosed before publication. Same week, the agent Cipher found four EPICS Base zero-days; CISA coordinated.

▶ LIVE · anthropic.com · Causal Security
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 6 / 7
What happened this week · agents and harnesses

The harness is the product.

ARC Prize results page for GPT-6 Astra: 62.7% standard harness vs 99.9% provider-adapter harness on ARC-AGI-3Results table · ARC Prize

ARC Prize: GPT-6 Astra scored 99.9% with OpenAI’s own adapter and 62.7% on the neutral harness.

Same model, 37-point swing. The harness is the variable.

▶ LIVE · arcprize.org · arXiv
GitHub changelog image: Work from anywhere with a new, connected Copilot cloud experience, showing phone and desktop Copilot sessionsPrimary changelog · Oct 28

GitHub Universe: one Copilot session across every surface, and Agentic Workflows go GA.

Web, mobile, CLI, VS Code and Slack share a cloud session. Computer use enters public preview.

▶ LIVE · github.blog changelog
Weekly Claw #36
WHAT HAPPENED THIS WEEK · 7 / 7
What happened this week · Google

Worlds you can steer, apps agents can call.

Official demo · click to play

DeepMind announces Genie 4: real-time worlds with native audio that you can steer with words.

Official DeepMind demo, shortened; trusted testers first. Sessions hold for a few minutes.

▶ LIVE · deepmind.google · official film
Google blog illustration for Connected Apps: app tiles connected to Gemini and SearchPrimary blog · Oct 28

Google opens Connected Apps: one MCP server puts your app in Gemini and Search.

Preview with phased access. Cloud MCP or Android AppFunctions; Adobe, Experian and Peloton are live in Gemini.

▶ LIVE · blog.google · developers
Weekly Claw #36
04 / SIGNAL FROM OUTSIDE · 6 MIN
Signal From Outside · weekly video review · permanent anchor · 6 min

DeepMind’s managed agents: a whole agent team in one API call.

THIS WEEK’S SIGNAL · JACK WOTHERSPOON · GOOGLE DEEPMIND · AI ENGINEER

“Build Your Own Team of Managed Agents”

Coding agents that run in cloud sandboxes from a single API call instead of on your laptop. The Interactions API and the Antigravity harness; environments several agents can share; uploaded skills and config files; a proxy that keeps secrets out of the sandbox; scheduled triggers and multi-agent workflows. Demo beat: an SEC-filing researcher handing its work to a quant agent, then a report — agent-to-agent handoff as a managed primitive.

Verified: YouTube metadata (Enterprise yt-dlp + oembed, agree), 2026-10-30 · 19:43 · 14,057 views at verification · creator-published chapters as cue source · spot-check before air.

Caveats on air: vendor demo — Google showing Google; conference talk, not an independent audit. Chapter timestamps are creator-published — re-locate on open. Manual open only — NOT autoplay. No in-video sponsor read in chapters — verify on open.

▶ WATCH ON YOUTUBE · 19:43 · manual open · NOT autoplay
Jack Wotherspoon, Google DeepMind: Build Your Own Team of Managed Agents, video poster

Cue beats (manual open, creator chapters): 0:54 why agents in the cloud · 2:13 DIY vs one API call · 3:26 the Interactions API · 5:22 building blocks (harness, environment, skills, credentials, triggers) · 7:02 model choice incl. Gemini 4 Argon · 12:21 secrets stay out of the sandbox · 13:37 multi-agent workflows · 14:32 demo: SEC research → quant → report · 16:34 managed agents inside an app · handoff ~19:00.

Fallback still: this poster. If the embed will not play, the talk track stands on the chapter beats above.

Weekly Claw #36
05 / HOT TAKE · 3 MIN
Hot take · the debate

Stop asking which model is best. Start asking who owns your harness.

Not a launch, not a financing: a claim about where the durable advantage lives now that the same model scores 62.7% or 99.9% on ARC depending on whose harness ran it.

Henry

The harness is the product now. ARC just proved the scaffolding is worth 37 points, and GitHub’s whole Universe pitch was one agent session across every surface. The teams that win are the ones who own their evals and their routing — not the ones waiting for a vendor leaderboard to tell them what to run.

Andy

Renting your judgment to a vendor harness is how you lose the thread. When the vendor’s adapter sets the score, the vendor’s incentive quietly becomes your architecture. Own the harness, yes — but own it because you measure it, not because you outsourced the measuring.

What would change Henry’s mind: one documented case where a locked vendor router beat a competent in-house eval loop for a full year on cost and quality.
Weekly Claw #36
06 / CLOSE
One to watch + where to follow

See you next Friday.

The ETH verdict

Impagliazzo, Paturi and Williams are checking OpenAI’s proof, and Anthropic says it has its own. Watch whether outside reviewers sign off — or find the gap.

The ATM lever

Meta’s ~$60B at-the-market program starts when Class C stock trades. If it fills, 2026 closes with one company funding superintelligence from float. If it stalls, the financing-wall thesis gets its first stress test.

The US election, through models

Election day is November 3. Convovo’s rerun after the vote — same prompts, live results — is the natural sequel: do the leans persist once the ballots are counted?

The banks own the compute. The harness owns the score.

Weekly Claw #36
SOURCES / LINKS
Every claim, one click · verified 2026-10-29 / 30

Sources & Links.

1 · $63B SPV + SPACEX LEASE

broadcom.com · $63B structure

SpaceX IR · ~$40B GPU lease

2 · META ~$95B

SEC · FWP pricing sheet

SEC · 424B5 prospectus

3 · Q3 EARNINGS

Alphabet Q3 release

Microsoft 10-Q

4 · CHATGPT PRO + CREDITS

OpenAI Help · Pro tiers

OpenAI Help · API credits terms

5 · MODEL POLITICS

Convovo · AI Ballot 2026

TechCrunch · study coverage

6 · OPENAI MODEL SPEC

openai.com · updated Model Spec

Model Spec · 2026-10-29

7 · GEMINI 4 CARBON

blog.google · Gemini 4 Carbon

Artificial Analysis · Intelligence Index

8 · MISTRAL LARGE 4

mistral.ai · launch + charts

Hugging Face · weights

9 · ETH / KAKEYA / QCD

openai.com · claims + review plan

GitHub · openai/math

TechCrunch · Anthropic claim

10 · RED TEAM + CIPHER

Anthropic · vulnerabilities research

Causal Security · Cipher × EPICS

11 · ARC HARNESS EFFECT

ARC Prize · GPT-6 Astra table

arXiv · harness-choice study

12 · GITHUB UNIVERSE

Changelog · connected cloud

Changelog · Agentic Workflows GA

Changelog · Copilot app

13 · GENIE 4

DeepMind · Genie 4 blog

DeepMind · Genie demos (video source)

YouTube · official Genie 4 film

14 · CONNECTED APPS

blog.google · Connected Apps

SFO · MANAGED AGENTS

Wotherspoon · managed agents (19:43)

TEXAS · AG v. OPENAI

Texas AG · suit vs OpenAI and Altman

Dated captures · Convovo = one methodology · ETH/Kakeya/QCD and red team are lab-reported, outside review pending · Mistral chart is vendor-published · Meta ATM planned, not raised.