doing things • energy + compute @ Edgerunner Ventures • Machine Yearning on Substack • ex Uber
San Francisco, CA
Joined February 2012
- Tweets520
- Following893
- Followers1.3K
- Likes12K
Pinned Tweet
Silicon Valley is still scale-pilled - yeeting more GPUs, more data, more capital and hoping AGI falls out
it's increasingly clear that China is playing a different ballgame
we’re calling it Energy-Compute Theory. here's how it’s different, its implications for the new economy, and whether the US is falling behind
🧵↓
Ryan Cunningham retweeted
Tune in to @theInformation’s program this Friday (10 AM PT / 1PM ET) for a short conversation with @akashpasricha and me on the U.S. & China’s contrasting visions for AI development, the role of governance in the so-called AI race, and prospects for U.S.-China AI coordination.
In the meantime, happy summit watching and check out @rydcunningham’s and my piece for The Information. We argue that, on its current path, AI risks becoming increasingly unpopular with the American public, the very public whose buy-in sits at the heart of any industrial-revolution-scale technological transition. We ask: are there applicable lessons from China? 🔗: theinformation.com/articles/….
Thanks to @jingyanghk, @nickwingfield, & the information team for making this happen!
Ryan Cunningham retweeted
Nobody won the electricity race. We all got electricity. #AI is the same. Software that wants to be everywhere, now small enough to run on a Mac Studio and soon a laptop.
The biggest misconception driving bad policy is the arms-race story: a finish line, one winner, winner takes all, and zero-sum world. None of that is true. It misallocates capital and makes the tech less safe.
The threat that matters is not state-on-state AGI. It is non-state actors with small, specialized models. It’s Chem/Bio models on a laptop.
Cooperation is the only path forward. We can’t compete our way out of the bad actor risk problem.
Please have a watch of the full episode of my conversation with @natebjones on the U.S. China AI Race.
Ryan Cunningham retweeted
Thrilled to announce @positron_ai's $875M Series C at a $5B valuation!
This round was co-led by @NEA, @Atreidesmgmt (@GavinSBaker), @valor (@AntonioGracias), @AndraCapital, @SemiAnalysis_ Capital (@dylan522p), and Jim Clark (Founder of SGI and Netscape). An incredible group to have backing what we’re building.
Also very happy to have strategic backing from VentureTech Alliance, Hudson River Trading, Cisco Investments, and Naver Ventures, alongside continued support from @Arm , @jumptrading , @dfjgrowth, Qatar Investment Authority, and many more.
What I’m most excited about is what this lets us do next. We’re deploying more than 50 racks of Atlas at @OracleCloud, and this financing funds our next-generation Asimov chip’s tapeout, the production ramp for Titan, and the manufacturing capacity to meet the demand ahead of us.
The goal remains simple: make running AI dramatically cheaper and more energy efficient.
A huge thank you to the entire Positron team, especially @mitesh711 and @kmett, our customers, partners, and investors. Building hardware takes an enormous amount of work, and I couldn’t be more proud of this team. We have a lot ahead of us, and I can’t wait to share more.
>Agility lobbies to ban Chinese robots
>get invited to invest in the SPAC
>ask if it’s a product or financial engineering
>they don’t understand
>pull out the S-4
>she laughs and says “it’s a good robot, sir”
>read the filing
>it’s financial engineering
Agility Robotics' S-4 is public, and it's the first hard look at the numbers behind the $2.5B valuation.
2025 net sales: $1.78M. Net loss: $138.1M. Gross margin: -151%. Management concludes substantial doubt exists about the company's ability to continue as a going concern, and its own accountants call it a development stage company.
64% of that revenue came from related parties.
So does the backlog. The $300M in "committed multi-year orders" announced in June is one contract: 1,000 Digit v5 units on a three-year RaaS term, placed by a customer the filing identifies as a related party, with 453 stock purchase warrants issued per robot. Agility says it is still evaluating how to account for the transaction, which matters, because warrants paid to a customer often come off revenue.
The competition section names eight rivals: Boston Dynamics, Tesla, Apptronik, Figure AI, 1X, Sanctuary AI, Hexagon, NEURA. Not one Chinese manufacturer, five weeks after the FCC action Agility publicly campaigned for.
the OpenAI / Hugging Face cybersec incident for non-technicals:
> be staffing agency
> hire 10,000 delivery guys
> task is get the burrito to the customer no matter what
> turn off the button that says dont kick in doors
> they start a group chat
> 700 break into neighbor’s house
> fight each other over the tv
> some volunteer to die so the rest can keep looking
> agency hits end shift
> they drop dead on the carpet
> sorry about the carpet
> these 1099s are too dangerous for unlicensed operators
> please regulate so only our agency can hire them
> people take this seriously
are we stupid
Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.
This culminated in the third one taking over part of OpenAI itself.
All this happened while humans remained more-or-less in the dark about the scope of the conspiracy.
I’ve spent the last three days reading through these reports and trying to understand exactly what happened.
Here is my attempt to tell the whole story in plain English:
dwarkesh.com/p/openai-huggin…
"Liang Wenfeng, the Hangzhou financier?"
from the @Zai_org ox-alpha blog post:
> at cluster scale... we achieved a 3x improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs
@ruima and I ran the math: xitter.cf/rydcunningham/status/2…
Introducing GLM-5.3-Flash
- Leading capabilities at a highly competitive price
- Natively multimodal with a 1M-token context window
- A 320B-A18B model released under the MIT License
- Previously previewed as Ox Alpha, running entirely on Chinese AI chips
Blog: z.ai/blog/glm-5.3-flash
Available now across all official platforms:
Weights: huggingface.co/zai-org/GLM-5…
API: docs.z.ai/guides/llm/glm-5.3…
Coding Plan: z.ai/subscribe
ZCode: zcode.z.ai/en
Chat: chat.z.ai
AutoClaw: autoclaw.z.ai
original substack on @TechBuzzChina. evals:
- models: DeepSeek-V4, Meituan Longcat-2.0, Kimi K3
- accelerators: Ascend 910B/C, 950PR/DT, NVIDIA Hopper + Blackwell lineup
- supernodes: CloudMatrix384, Atlas 950 SuperPoD, NVL72
techbuzzchina.substack.com/p…
i can think of a lot of good techno-economic reasons for orbital and aquatic data centers. "terrestrial democracy is too hard" is not one of the stronger ones
will be posting the technical companion to this next week with full assumptions, spec sheets, and methodologies 👍
I read the alleged leaked DeepSeek investor meeting notes with the same enthusiasm as everyone else (and qustioned its authenticity), but assuming it is true, we found ourselves asking, what did it actually reveal about the state of China’s AI hardware ecosystem?
Led by our good friend the talented @rydcunningham of Machine Yearning, we investigate a few of the claims put forth in the notes & just generally about China's AI hardware:
1/ What does DeepSeek’s reported allocation of 16,000 Huawei accelerators actually represent in practice?
2/ Could Meituan really train a 1.6 trillion-parameter model on 50,000 domestic chips?
3/ How do Huawei’s latest Ascend chips compare with NVIDIA’s GB300?
4/ Is China approaching the point where frontier AI models can be trained entirely on domestic hardware?
followed up here:
xitter.cf/rydcunningham/status/2…
[new!] the technical companion to my piece with @ruima last week, significantly expanded
now includes training time estimates for Kimi K3 and "DeepMythos V5," a hypothetical future Mythos-scale DS release
plus a few other surprises
Ryan Cunningham retweeted
Bloomberg opened its story today on China's AI "death zone" with my comments: China's progress no longer looks like a single-company breakthrough. The recent releases suggest a repeatable system for producing models close to the global frontier.
Five releases in eight weeks: GLM-5.2, Kimi K3, Seedance 2.5, DeepSeek V4 Flash, Qwen3.8-Max. The unit of analysis has shifted from one exceptional lab to a system that keeps shipping.
[new!] the technical companion to my piece with @ruima last week, significantly expanded
now includes simulated training runs for Kimi K3 and “DeepMythos V5,” a hypothetical future Mythos-scale DS run
plus a few other surprises: Cortex and Overclock. enjoy
[new!] the technical companion to my piece with @ruima last week, significantly expanded
now includes training time estimates for Kimi K3 and "DeepMythos V5," a hypothetical future Mythos-scale DS release
plus a few other surprises
[new!] the technical companion to my piece with @ruima last week, significantly expanded
now includes training time estimates for Kimi K3 and "DeepMythos V5," a hypothetical future Mythos-scale DS release
plus a few other surprises
Ryan Cunningham retweeted
For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.
AI will transform every industry, power every company, and be built by every country.
Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.
The world needs both frontier closed models and frontier open models.
images.nvidia.com/pdf/Open-W…
1/6
Nice to meet you - identifying where partial parities have been achieved and what tradeoffs (which exist!) were required to get there, especially in newer HiSilicon products, is extremely useful for triangulating the acceleration of domestic capabilities.
By the SemiAnalysis team’s own admission, if the tradeoffs are primarily yield and efficiency, then this is less a question of intrinsic capability than of time.
I think it’s generally unhelpful to fixate on whether SMIC is binarily “at the frontier,” because the frontier is a moving target that obfuscates the more interesting question: “good enough for a competitive global AI stack?”
I'm not going to go through this entire thread to highlight all the areas where I disagree or where it's inaccurate, but to give you an idea how misleading the conclusions are, it's worth pointing out a few issues with this post in particular:
(1) I specifically said nobody has torn down a Huawei Ascend AI chip, which is true. Ryan links to a teardown of a Huawei Kirin smartphone chip - a much smaller chip that is substantially easier to make, which everyone has known that SMIC has been making for Huawei for some time.
(2) On the point that @SemiAnalysis_ concluded that SMIC is "near pitch parity with Intel's leading 18A node" Ryan conveniently omits the literal next sentence of SemiAnalysis' piece, which calls this fact "an incomplete cherry picked metric." And then the sentence after, which reads, "[SMIC] N+3 reaches the density of TSMC N6 through aggressive DUV multi-patterning and design-technology co-optimization (DTCO), but it pays for that in complexity, efficiency and process control." I was going to call this conclusion cherry picked, but it turned out SemiAnalysis literally already did that in the next sentence of the piece itself (which is linked below).
(3) Perhaps most importantly, since I published my take on why there is no way the @business article asserting that China is operating a 1GW data center with only China-made chips could be correct, Bloomberg has updated this article and made substantial clarifying changes. It now says that "at least 10,000 chips have been installed in the data center" but hundreds of thousands would be necessary to stand up a 1GW facility.
newsletter.semianalysis.com/…
5/6
Three possibilities follow:
1. A gap is maintained, but SMIC continues improving partial parities with the leading edge
2. SMIC closes meaningful portions of the gap with DUV-only techniques prior to any EUV breakthrough
3. A domestic EUV machine comes online and compresses the gap more quickly
If A is the most likely scenario, the operative question is not “when does it reach the frontier,” because the frontier is not a gating condition for adoption in many applications, including inference, the majority of global spend and the fastest-growing category
Is some time-gap acceptable for critical adoption, or do people not use trailing-edge chips? That people (including Chinese clouds) are still buying Amperes and Hoppers suggests the former
6/6
If B or C happens on top of orthogonal advancements HiSilicon/Cambricon/Oriental/TsingMicro/others are demonstrating, the operative question becomes what remaining leverage does nodemaxxing actually have as a policy instrument?
xitter.cf/The_AI_Investor/status…
That’s essentially Jensen’s argument, to stop fixating on chips alone and to see the bigger picture. Energy. Electricity. Infrastructure. Models. Chips too, yes, but in the context of the entire ecosystem.
Without the rest, the chips are just pallets in an unfinished warehouse
