Cyber defense requires reliable reasoning across noisy logs, events and network flows.
@CormaAI is training foundation models for exactly that. They've increased server capacity 24x to meet demand.
Huge congrats to @AlonPluda and team on the $125M Series A!
Fastest angel check I ever wrote. Proud to back them at @khoslaventures too!
We raised a $125M Series A at a $1B valuation, 3 months out of stealth.
@NEA led. @sequoia, @khoslaventures and @coatuemgmt doubled down.
@CormaAI is training foundation models toward superintelligence for defensive cybersecurity.
Because the greatest gift cyberattackers have ever received is smarter AI, and right now only one side is armed.
Arm the defenders.
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
incredible work from will
i don’t wanna hear the ‘well duh’ crowd.
"It's not what you know, Clyde. It's what you can prove in court."
i passed this onto someone close with trump and he’ll make a public comment tomorrow.
due to the implications. this may go down as one of the most important pieces of research in human history.
I have a major result to share.
I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China.
This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model.
Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves.
I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems.
I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used.
It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20.
These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs.
I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo….
There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
Top 30 Startups and BigTech companies based purely on ~50,000 engineer job changes in the last 3 years.
If engineers leave a company to join another, it ranks higher especially if the company they come from ranks highly.
Here are the final results:
Having led infra at Citadel Securities and OpenAI, I know how hard it is to understand why complex production systems fail.
The amazing @traversal_ai team is tackling this challenge with AI at enterprise scale, and I was blown away by how well their system works.
Proud that we @khoslaventures co-led their $120M Series B! Huge congrats @_anish_agarwal and team!
LFG!! 🚀
We just raised $120M, with a $1B valuation, to build software that keeps itself running.
AI is writing more code than ever. It’s also making production less reliable: more code, shipped faster, understood by few.
@traversal_ai understands how production systems fit together, finds the root cause of incidents, and helps prevent the next one.
Engineers at Nubank, Harvey, PepsiCo, DigitalOcean, and a growing list of Fortune 100 companies use Traversal every day.
Here’s production without the drama 👇
.@bobmcgrewai is best leader I've ever worked for. Also, he's got impeccable taste in picking industries and companies (PayPal > Palantir > OpenAI).
Huge congrats to @bobmcgrewai and the whole Arda team. Proud to be an angel investor, and proud that @khoslaventures partners with Arda, too!
We started Arda in 2025 to speed up American manufacturing. We have spent the last year building with Anduril as our design partner. Today we are launching the company publicly on CNBC and announcing our Series A: $160M at a $1.5B valuation.
Our platform helps manufacturers design and operate manual assembly lines, putting designs into production faster now. Next we'll build automated lines that can be reconfigured as quickly as human workers, letting America make products that now require armies of workers in Asia.
Satya is right. The way to make SI safe is not to train it with a sense of self, its own moral philosophy, and permission to act as a conscientious objector. That’s the “alignment” approach and it magnifies the control problem.
The engineering approach that Satya describes is different: separate the supply of intelligence from authority over it. Surround non-deterministic models with deterministic controls, observability, privilege limits, logging, and the ability to always contain or shut them down. Treat models/agents like powerful insider risks, not moral patients whose psychological wellbeing is at stake.
As Satya points out, the most trustworthy system is the one that lets us trust the model the least — not the one that encourages the model to develop independent agency and grievances. Engineering safety is not the same thing as “alignment.”
We @khoslaventures are proud to work with @arena, led by the amazing @ml_angelopoulos. They built the industry's most important AI leaderboard, and are now measuring how agents behave in the real world. This work is critical to delivering safe & aligned AGI!
Today, we're announcing our Series B: $200M at a $3.1B valuation, and the release of Arena's Alignment Index.
AI is advancing faster than our ability to evaluate it. The world needs a neutral third party to measure how safe and aligned AI actually is once it's in the hands of real people. That's the role Arena is stepping into today.
The Alignment Index is built from real-world agent traces, starting with three signals: Unauthorized Action, False Attribution, and Deceptive Completion, with results for over 20 frontier models launching today.
Hear more from our CEO @ml_angelopoulos below, and read the full breakdown, and more details about our company growth in thread.
This funding will allow us to go further on safety signals, agentic capabilities, and modalities. It will also allow us to grow the company to match the scale of our mission. Today, we’re still a nimble team of 90, and we're looking for people to join us across research, product, engineering, and more. Come build with us.
As AI gets more capable and autonomous, understanding how well it can safely and effectively work for people only matters more. That's what Arena is built to measure and advance.
Thank you to our community, our investors, and everyone building with us!
Take a look at our open roles if you want to join us (link in bio).
youtube.com/watch?v=8s7uYtCM…
Big day today. Our interview with Richard Ho (@rh00), the VP Hardware at OpenAI, is now live!
We discuss the new chip, chip design, how OpenAI views the ecosystem, and what's coming down the pipe. Very insightful, thanks Richard!
yeah this checks out
my first few months at periodic, the weekly midtraining meeting was:
1. guy who trained the first trillion param LLM
2. guy who invented the attention mechanism
3. me
you learn very quickly surrounded by this density of talent. we're hiring btw.
Thrilled to join @khoslaventures and learn from @vkhosla, @SStrohband, @SamirKaul1, @rabois and David, and from all of my amazing new partners, while helping back and build the next generation of world-changing companies.
Someone asked me yesterday what success looks like 5 years from now. I think it’ll probably be too early to judge my investment record. But I hope founders will say I was one of the most helpful people around the table, and had a real positive impact on their company.
If you’re building in AI, especially AI infra, I’d love to talk! 🚀
Thrilled to announce @TalBroda has joined Khosla Ventures! He’s bringing his experience building AI infrastructure to the founders we back.
At OpenAI, he led more than a dozen teams that built and operated the hardware and software serving OpenAI’s models. He helped lead GPT-4.5 training, participated in the development of OpenAI’s first custom inference chip, Jalapeño, and contributed to Stargate. Previously, he headed up platform at Citadel Securities. He’s tackled the infrastructure challenges AI founders face firsthand, and we’re excited to have him in join KV!
wsj.com/pro/venture-capital/…
"After 21 hours spent searching this data by roughly 950 agents using 210 million tokens, one of the agents spotted something remarkable..."
The bitter lesson is everywhere.
If you scale search and compute enough, you start finding things we humans missed.
Biology is no exception!
Claude has discovered a previously unknown enzyme system hidden in the DNA of bacteriophages. Beside the enzyme’s gene sits a long array of repeating DNA—a structure that looks somewhat similar to CRISPR.
We don’t yet understand what this system does, but only a handful of known systems share its features, and all of them are able to cut, copy, and paste DNA. Historically, the discovery of such programmable systems has helped revolutionize medicine. CRISPR, for instance, is now the foundation of genetic medicines. But it will take much more work to learn what this system does, and whether it can be put to similar use.
Read more: anthropic.com/news/claude-di…
.@periodiclabs is an insanely talented team, tackling some of the hardest problems in science, and delivering awe-inspiring progress at the frontier of AI and physics at remarkable speed. Read about their work below!
We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next.
Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon.
This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials.
Read our blog posts below.