Episode 27 · Friday, August 28, 2026 · 4:00 PM ET

Weekly Claw

The agent owns the loop.
Nvidia moved to buy the open-weights commons; Zhipu unmasked Ox Alpha as GLM-5.3-Flash serving 100T tokens a day on Chinese chips; OpenAI stacked chip, model, seat and said AGI by December; a four-month-old assistant raised at $2.5B. Eight stories, less time on each, more of the week.

Hosts: @AndyML · @HiM ~40 min · nine stories · faster cuts · built to be clipped

Sponsored by Heritage Telecom Herald Labs

“Who runs the loop, and who decides what it does next.”

Weekly Claw #27

01 / COLD OPEN

Cold open · the frame

The agent owns the loop.

Eight stories, one frame: the consolidation race. Chips buying commons, stealth models on domestic silicon, four-month-old agents at $2.5B, AGI dates with footnotes. The moat is no longer the model. It is the operating layer.

1

Consolidation

2

Sovereign compute

3

Agent capital

4

AGI claims

5

Operating layer

🌱Nvidia is buying Hugging Face for $12.9B; the GPU vendor would own the open-weights commons.

🐴Ox Alpha was GLM-5.3-Flash: 100 trillion tokens a day served on 100% Chinese chips.

💰Altman says OpenAI has AGI internally by year-end; Mark Chen calls it 80% of the way there.

Weekly Claw #27

WHAT HAPPENED THIS WEEK

What happened this week · part 1

Consolidation, sovereign compute, agent capital.

Four moves that defined the week. Each card links to the primary source.

Nvidia is buying Hugging Face.

$12.9B reported, unconfirmed. The GPU vendor would own the model commons.

cnbc.com

Ox Alpha is GLM-5.3-Flash.

Serving ~100T tokens/day on ~100K Chinese GPUs (Zhipu-reported). Benchmarks over press.

benchmarks (x.com)

OpenAI Jalapeno vs Nvidia Blackwell analysis

OpenAI ships Jalapeño.

First custom inference chip; vendor-reported wins vs Blackwell on perf/watt.

semianalysis

OpenAI says AGI by December.

Altman in TIME: internal system he'd call AGI by end of 2026 — his definition.

time.com

Weekly Claw #27

WHAT HAPPENED THIS WEEK

What happened this week · part 2

The operating layer ships.

Agent capital, open weights, always-on agents, local-first stacks — the deployment story.

Qwen opens Flash-Next.

Qwen4-preview architecture ungated: 125B params, 6B active, 250K context.

qwen.ai

Laude ships Headlong.

The agent that never sleeps — fixed a recall bug in 48 min, no human in the loop.

laude.org

Perplexity puts the stack on the desk.

Portable Computer runs model, harness, sandbox locally on DGX Spark.

perplexity.ai

Instinct raised at $2.5B.

$250M new at a $2.5B valuation, four months old. Agent capital is hot.

techcrunch

Weekly Claw #27

SPONSOR

Brought to you by

Heritage Telecom keeps the lights on while we keep the operating layer honest. Independent infrastructure for independent voices.

Independent. Reliable. Quietly essential.

heritagetel.com

Weekly Claw #27

03 / SIGNAL FROM OUTSIDE · 7 MIN

Signal From Outside · weekly video review · permanent anchor · 7 min

When agents contribute, everything breaks.

THIS WEEK’S VIDEO · GITHUB

“OpenClaw Went Viral. Meet the Maintainers Building and Securing It.”

Four maintainers on the couch · 46:00 · 388K stars, 80K+ commits, ~100K issues and PRs from 2,000+ contributors in nine months. The first codebase at full agent-contributor saturation: prompt requests, badge gaming, edited-not-bounced PRs, agent review, zero telemetry. A field report from the front, not OpenClaw news.

▶ PLAY · youtu.be/5VSwaUXtPIE

OpenClaw Went Viral — Meet the Maintainers Building and Securing It (GitHub)

Weekly Claw #27

04 / HOT TAKE · 3 MIN

Hot take · the fight · two sides, one verdict

Nvidia buying Hugging Face is bad for open source.

The motion. Twelve point nine billion dollars, still unconfirmed, for the hub where the open-weights world ships. Henry argues it is bad for open source. Andy argues it is good. Two minutes each, then the verdict.

HENRY · BAD FOR OPEN SOURCE

The commons gets a landlord.

The biggest GPU vendor now owns the hub its own competitors ship on. Every neutral surface — model hosting, benchmarks, Spaces — now has an owner with a dog in the fight.

Conflict of interest: the referee sells jerseys
Microsoft-GitHub precedent: openness survives, neutrality does not
The rent: terms of service beat licenses when the hub is single

ANDY · GOOD FOR OPEN SOURCE

Validation, not capture.

Nobody pays twelve point nine billion for a charity. Open source just got priced as the strategic layer of AI — and Nvidia sells GPUs to everyone, so walling the garden shrinks its own market.

Capital: the commons stops living donation to donation
Incentive: locking HF hurts Nvidia’s own TAM
Fork discipline: open weights cannot be unshipped

"The question is not whether the hub stays open. It is who sets the rent — and that is why your right to re-deploy is the moat that is still open."

qwen.ai/blog?id=qwen3.8-flash-next laude.org/updates/headlong-a-microharness-for-persistent-agents perplexity.ai hub (local-first agent)

Weekly Claw #27

SPONSOR

Also brought to you by

An applied AI product lab where humans and agents build together. Entity is mission control for agent teams. Hacker houses worldwide.

labs.theherald.co

Weekly Claw #27

05 / CLOSE

One to watch + where to follow

See you next Friday.

The vertical stack

OpenAI's Premium number is the proof point. Five times Standard usage, no five-hour cap. If that lands, every other inference vendor has to publish a five-times-or-better response, and the model layer stops pricing by tokens.

The deployment layer

Qwen3.8-Flash-Next day-zero stability on SGLang will determine whether the open-weight ecosystem actually wins a quarter. The hosted QwenCloud API is still “coming soon.” Independent evals are the next test. Figure's Index — 16M videos, $15M paid to creators — doubles down on robots as a data business.

The watts

Perplexity Portable Computer needs a hardware-cost receipt. “Zero token cost” is a slogan until someone publishes the watts, the cooling, and the time-to-first-token on a real desk.

Follow the operating layer.

Weekly Claw #27

SOURCES / LINKS

Every claim, one click · verified 2026-08-28

Sources & Links.

S1 · NVIDIA × HUGGING FACE

cnbc.com: Nvidia buys Hugging Face $12.9B arstechnica.com: Nvidia to acquire Hugging Face

S2 · OPENAI STACK

openai.com/index/jalapeno-first-results/ openai.com/index/premium-seats-chatgpt-business/ help.openai.com ChatGPT Business overview NVIDIA investor: Groq 3 LPX in production

S3 · QWEN FLASH-NEXT

qwen.ai/blog?id=qwen3.8-flash-next huggingface.co/Qwen/Qwen3.8-Flash-Next (BF16) huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8 docs.sglang.io cookbook: Qwen3.8-Flash-Next

S4 · HEADLONG

laude.org/updates/headlong github.com/laude-institute/headlong (Apache 2.0) @andykonwinski launch post

S5 · PERPLEXITY PORTABLE COMPUTER

perplexity.ai/hub/products/portable-computer perplexity.ai hub: local-first agent VentureBeat launch coverage (Aug 25)

S6 · FIGURE INDEX S7 · GLM-5.3-FLASH / OX ALPHA

scmp.com: Ox Alpha = GLM-5.3-Flash wccftech.com: GLM-5.3-Flash on Chinese GPUs

S8 · INSTINCT RAISE

techcrunch.com: Instinct raises at $2.5B valuation

S9 · OPENAI AGI CLAIM

the-decoder.com: Altman AGI claim felloai.com: what Altman actually promised

figure.ai/news/introducing-index Ars Technica: World Humanoid Robot Games Euronews: Beijing robot games video

SIGNAL FROM OUTSIDE

x.com/davidfromkansas Codex virtual-office clip

HOT TAKE · DEPLOYMENT LAYER

qwen.ai Qwen3.8-Flash-Next release laude.org Headlong launch

Vendor-reported claims are labeled on each slide (OpenAI work-per-watt by published TDP; Qwen benchmark vendor-run; Perplexity 82.6% vs 74.0% self-authored; Figure scale company-reported; Nvidia×HF price press-reported, unconfirmed; GLM-5.3-Flash scale Zhipu-reported; AGI date is Altman’s own definition, not independently validated). The hosted QwenCloud API is “coming soon.”