Steven Xi retweeted
Nvidia’s board stated today that it has found “no reason to believe that either national security or Nvidia secrets were compromised” as a result of Jensen’s actions.
They are initiating an independent investigation overseen by a committee of independent directors, assisted by outside independent counsel, that will examine all available information. This will be an unbiased, exhaustive and thorough investigation given the fiduciary responsibilities of these independent directors.
Jensen said that “I support the investigation and will cooperate fully. I love our country and our company. I am committed to protecting NVIDIA and supporting America’s national security. I am proud and grateful to be an American.”
I have known Jensen for almost 30 years. I simply do not believe he would ever knowingly do anything to disadvantage either America or Nvidia.
The only thing he has ever said to me about China is the absolute importance of keeping the American hardware ecosystem, led by Nvidia, ahead of China. That’s it.
More detailed thoughts follow below:
They are initiating an independent investigation overseen by a committee of independent directors, assisted by outside independent counsel, that will examine all available information. This will be an unbiased, exhaustive and thorough investigation given the fiduciary responsibilities of these independent directors.
Jensen said that “I support the investigation and will cooperate fully. I love our country and our company. I am committed to protecting NVIDIA and supporting America’s national security. I am proud and grateful to be an American.”
I have known Jensen for almost 30 years. I simply do not believe he would ever knowingly do anything to disadvantage either America or Nvidia.
The only thing he has ever said to me about China is the absolute importance of keeping the American hardware ecosystem, led by Nvidia, ahead of China. That’s it.
More detailed thoughts follow below:
213
190
2,159
596,175
Love this interview. One of the best interviews given by @bernhardsson, who really came through as a thoughtful and authentic founder with sense of humor and humility. I bet founders and developers want to hear more from Erik. Kudos to @HarryStebbings too for a job well done!
Spotify 👉 open.spotify.com/episode/6DDm5l…
Youtube 👉 youtu.be/1dwyL6Iz_PU
Apple Podcasts 👉 podcasts.apple.com/us/podcast/20v…
2
0
4
170
Well said bout Databricks evolution from open source Spark to today’s power house in the data layer for AI!
The reason Databricks is the largest open source company in history is both structural and execution focused. Ben gives a great overview but as an open source founder I’ll try my hand as well.
Databricks is at its core a compute company. Their project Spark beat out all the other distributed data processing architectures like Hadoop and Cassandra to emerge victorious in the big data wars of the 2010s.
Partially this was because the other architectures were flawed and partially because they invested more in fixing them through hiring incredible engineers. This company did come out of a Berkeley PhD lab.
Still they struggled to make money. So in the mid 2010s they began making their next investment in a proprietary execution engine for Spark they called Photon.
As if building the leading distributing data processing library wasn’t hard enough they also shipped a native C++ vectorized query engine and continued both the public Spark roadmap and an internal proprietary roadmap as well.
It’s worth pointing out that this company is a decent example of the a16z thesis that you can kingmake with capital. Ben dropped a $10m check on a company that only asked for $250k, which is how the company had the war chest to make these investments.
If that’s all they had done, my guess is they’d be a $20B or $30B company today. I’ve made the point before that open source infrastructure expands forever.
That said there was a big second act. Circa 2018 they pioneered lakehouse architectures. The way to think about it that the traditional way you provision databases is you reserve some bucket on specialized compute and storage for them on your infrastructure. This made sense in a pre-cloud world where you had separate special-purpose servers, but less so in a world where you have scalable services like S3, EC2 etc.
So instead of reserving expensive DB compute and storage you reserve cheap S3 storage and compute that can spin up and down based on variable usage.
And now there’s a third act, namely building a cloud with AI primitives. The new gateway is a big piece of that but last month they also released a backend as a service with the primitives for building agents.
So again, why is it such a big company? Again, I think this is both structural and execution oriented.
The structural reasons: they started in an extremely lucrative market with a barrier to entry (it is very difficult to create popular open source projects in valuable markets!), that expands forever.
Second, they are a general purpose orchestration framework.
What I mean here is that every open source / devtools market is split between general purpose orchestration and specialized single purpose utilities. Think auth (Clerk) or email (Resend) for web apps (Next.js), or integrations (Composio) and sandboxes for agents (Mastra). While it is simpler from a product execution standpoint to build a single purpose utility, it is usually more valuable to build the general purpose orchestration framework and Databricks bears that out when you compare them to single purpose utilities doing things like ingestion and transformation (dbt, Airbyte).
Lastly as an orchestration framework / compute company, unlike database companies they are not limited to a particular workload (Supabase etc for transactional data, Clickhouse etc for columnar data), and they are in the core infra layer as opposed to being stuck on the client (Vercel/Next.js). It’s the $10B vs $200b difference
The execution reasons: they are technical geniuses who really want to win. I’ve spent the last decade building open source companies (Gatsby and now Mastra) and the typical open source founders are brilliant tinkerers and, in the best case, it takes a while to develop a ruthlessly efficient competitive commercial DNA. With Databricks they had both sets of DNA within the founding team — incredibly rare.
Congrats to @alighodsi and team on everything they’ve built so far!
2
0
4
384
Steven Xi retweeted
People will eventually realize that inference providers can't win on price alone.
Data security and reliability will matter more as competition intensifies. Startups racing to the bottom have little chance of surviving.
Yes, small inference teams might still make money, but once you factor in the years they've invested, the returns may hardly justify the effort.
Data security and reliability will matter more as competition intensifies. Startups racing to the bottom have little chance of surviving.
Yes, small inference teams might still make money, but once you factor in the years they've invested, the returns may hardly justify the effort.
3
1
11
1,423
Steven Xi retweeted
Databricks Chief AI Scientist Jonathan Frankle (@jefrankle) says the bottleneck to enterprise AI adoption is evals, not models.
"The models are good enough at a lot of really complex enterprise tasks and knowledge tasks"
"the real challenge is describing what you want AI to do did it"
"The models are good enough at a lot of really complex enterprise tasks and knowledge tasks"
"the real challenge is describing what you want AI to do did it"
27
21
193
28,462
Cool cap! I wear it all the time, and developers often stop me because they assume I work at @Modal. I always ask what they like about it, and the answer is overwhelmingly positive. Developers love it!
0
0
2
197
Steven Xi retweeted
OpenAI and Anthropic will be the two best stocks to own in history.
Diffusion models, VLAs, World Action Models, and other specialized architectures will still have their place, but they'll all end up as tools orchestrated by one single powerful LLM-based reasoning models.
The same goes for agent harnesses. Foundation model companies will eventually own the general-purpose agent harness layer, leaving little room for independent players outside of specialized vertical use cases.
Diffusion models, VLAs, World Action Models, and other specialized architectures will still have their place, but they'll all end up as tools orchestrated by one single powerful LLM-based reasoning models.
The same goes for agent harnesses. Foundation model companies will eventually own the general-purpose agent harness layer, leaving little room for independent players outside of specialized vertical use cases.
1
1
10
1,140
Steven Xi retweeted
I interviewed @neilmovva for @substratemag on latency-insensitive inference, alternative chips, and why massive datacenters face diseconomies of scale.
"We've always predicted that async inference is going to eat the world. That could easily be 95% or more of inference demand."
ntier of inference for years, from NVIDIA to co-founding @sailresearchco. I learned a lot from our conversation - you can listen below.
"We've always predicted that async inference is going to eat the world. That could easily be 95% or more of inference demand."
ntier of inference for years, from NVIDIA to co-founding @sailresearchco. I learned a lot from our conversation - you can listen below.
34
25
340
39,789
This is no surprise to us at @eastlinkcap! LFG!🚀
The first time we talked to @bernhardsson in 2023, we knew he was someone special and would be an exceptional founder. Proud to have backed them with high conviction since the Series A. Congrats @bernhardsson, @akshat_b and team on $1B run rate and the Series D! We believe @modal will be one of the most important AI Infra companies. We're grateful and look forward to continuing to be a small part of the journey 🚀
0
0
1
152
The first time I spoke with @bernhardsson in 2023, I was immediately sold on his vision. As a fellow physics major, I’ve always appreciated his first-principles approach to problem solving. We’re grateful to continue being a small part of the journey!
0
0
4
220
Well said. I offer a similar advice to new graduates (except PhDs) including my own kids: Cursor and Mercor are the exceptions rather than the rules.
I work at a fast-growing startup that I am very optimistic about.
And I still think joining a startup as a new grad is a high-risk career decision.
Not because startups are bad. A great startup can compress years of learning into months. You can own real problems, talk directly to users, make decisions far above your nominal level, and watch your work affect the company immediately.
But the downside is often misunderstood.
The obvious risk is that the company fails and your equity becomes worthless.
The more important risk, especially early in your career, is that your learning environment fails with it.
Your first few years of work are when you build your internal definition of what good engineering, good management, good product judgment, and good execution look like.
At a mature company, many of these systems already exist. There are imperfect but established review processes, senior engineers, historical context, clear ownership, customers, metrics, and people who have seen the same class of failure before.
At an early startup, you may be learning while the organization itself is still learning how to operate.
That can be amazing if the people around you are exceptional.
It can also be dangerous.
A company can call something ownership when there is simply nobody available to mentor you.
You can report directly to a founder and still receive very little useful feedback.
You can touch strategy, product, sales, and engineering and feel that you are growing incredibly fast, while never becoming excellent at any of them.
You can become indispensable because only you understand a broken process. That makes you important to the company, but not necessarily more capable.
And because everything changes so quickly, it is easy to mistake surviving chaos for learning.
I have seen friends struggle with exactly this.
Smart people joined exciting companies, worked extremely hard, wore five hats, and after a year or two had trouble answering a surprisingly simple question:
What am I now much better at than when I joined?
That is the question I would care about most.
For a new grad, I would evaluate a startup less by how impressive the founders sound and more by the quality of the learning loop:
Who will review my work?
Who can tell me when my reasoning is wrong?
Will I own something end to end?
Will I see real users and real outcomes?
Can the team explain what good performance looks like after 3, 6, and 12 months?
If the company disappears in two years, what skills, judgment, and evidence of work will I carry with me?
And of course: does the company actually have enough runway for me to stay long enough to learn any of this?
The best startup jobs give you unusually tight feedback between action and consequence. You build something, ship it, watch it fail, talk to the user, fix it, and gradually develop judgment.
That loop is incredibly valuable.
But without good mentorship, clear ownership, and real feedback, the same speed can simply generate noise.
So yes, I am very bullish on startups.
I am also cautious when people tell new grads that joining one is simply a matter of being brave enough to take risk.
Risk is only worth taking when it buys you something.
For an early-career engineer, the most valuable return is not a title, proximity to a founder, or even equity.
It is becoming meaningfully better at something that will still matter after the company is gone.
1
0
7
626
Thanks for sharing, chenyang! Agreed that agent Infra is urgently and badly needed!
My biggest takeaway from SF Tech Week: the center of gravity has already moved.
Last year everyone was talking about agents.
This year, a lot of the interesting conversations were about everything agents need once they actually start working: routing, search, data, evals, observability, latency, inference cost, GPUs, deployment, reliability.
The question used to be:
Can an agent do this?
Now it increasingly feels like:
How do we make a million agents do this reliably, cheaply, and controllably?
That shift matters.
Companies like Tavily are a good example. Most end users may never know the name, but a huge number of agent workflows need search and web retrieval underneath. Once agents become real products, the boring-looking infrastructure around them becomes extremely valuable.
Another thing Tech Week messes with is your perception of startup outcomes.
You meet founder after founder who raised, got acquired, or somehow went from starting a company to becoming a VP / GM / Head of X at a much larger company in a few years.
After hearing enough of these stories, your brain starts treating this as normal.
It obviously isn't.
Tech Week is basically the Instagram feed of the startup distribution. You are meeting people from the far right tail. The companies that ran out of money, never found PMF, or quietly dissolved are not usually on stage telling you their story.
And even we got acquired can mean very different things: a great strategic outcome, an acquihire, investors getting some money back, or simply a soft landing.
Startups don't just increase expected upside. They massively increase variance.
One other thing I kept hearing was FDE: Forward Deployed Engineer.
The role is interesting because it blurs boundaries that used to feel much cleaner.
Engineer builds.
Sales sells.
PM translates.
But more and more, the valuable person seems to be someone who can talk to the customer, understand the business problem, talk to engineers, turn ambiguity into something executable, and actually ship it.
Engineer + Product + Consultant + Sales.
I increasingly think this is less a job title than a survival skill.
AI is already crawling into every industry and every workflow. So maybe the interesting question is no longer what AI can do.
It is:
What workflow has not been rebuilt yet?
And after spending a few days around all of this, I came away with two contradictory feelings:
How is everyone here doing something this ambitious?
And:
Maybe I should be doing more too.
The Bay Area has never had a shortage of ideas or smart people.
The scarce things are judgment about what is worth doing, the willingness to stay with it for years, and the ability to get other people to care.
0
0
1
95
Bold prediction: Runtime will be one of the coolest tech conferences in SF, on par with Databricks’, in two years! Feel so blessed to be a small part of the journey early with GOAT @bernhardsson and his team!
We threw our first conference last week in SF. Tons of cool people and talks. A very high correlation between being cool and being there.
Anyway, this is the keynote, in which I announced a bunch of Modal features. We also launched 1M sandboxes in a minute for fun!
0
0
1
140
Thank you, Moin! It is always a pleasure to hear you are happy camper with our friends at @modal! 🚀
Even in moments of stress, it’s so wonderful to work with the @modal team. They respond quickly, do what they say they’re going to do, and it feels like you’re talking to a real engineer rather than a support person.
0
0
0
103
Good insights from our friend @ttunguz! Agreed with @buckymoore, and glad to report that we also backed @togethercompute and invested in @modal way back in 2023. We believe AI Inference is the biggest market for AI, even larger than coding.
Earlier this year, when asked by @tbpn what the most underrated market in venture was at the time, I said "inference." This is also why I have proudly backed companies like @togethercompute @inferact @reactorworld over the past few years.
Per usual, @ttunguz provides a fantastic articulation of the why behind this statement.
0
0
0
188
Well said, @quxiaoyin! We are lucky to have you working with us and share your always sharp insights!
We are in the fierce inference competition phase.
To make money, owning GPU is not enough.
You must:
- know where to get cheap electricity (even: Finland, Philippines??)
- ideally own GPU, not rent GPU
- good at running inference optimization
- sometimes even post train your version of the model to make it faster and efficient
The inference market is getting more sophisticated. We used to say DeepSeek is cheap based on their listing price; now you can easily find heavily discounted token price off their already low listing price, from many inference providers.
That’s being said, openrouter is a fucking huge business. I almost feel like they shouldn’t sell. Because on top of many labs, you also have way more inference providers for each model competing fiercely.
Funny that people used to think openrouter is a niche business and Claude was gonna own all models just a few months ago.
1
0
1
357
Love this partnership: It gives enterprise customers what they need the most: context and data intelligence! Smart move, Reynold @rxin!
Excited to announce our new partnership with @DecagonAI to bring richer enterprise data intelligence directly into customer conversations.
AI agents are only as useful as the context they can access.
Enterprises already have an incredibly rich record of their customers in Databricks. Orders, payments, product usage, preferences, and countless other signals all help explain who a customer is and what they need.
Through this partnership, Decagon agents can access governed data from the Databricks when they need it, while insights from those conversations can flow back into Databricks and become useful across the rest of the business.
We’re excited to work with the Decagon team to make enterprise customer agents more context aware while keeping data governance at the center.
decagon.ai/blog/decagon-a…
0
0
0
122
Steven Xi retweeted
HUGE days of product launches for @modal!
Just launched at Modal Runtime:
- Multi-node clusters
- VM Sandboxes
- Sandbox Sidecars
- Sticky Sessions
And more:
2
1
25
1,677
This @modal guy is a rock star! Spinning a million sandboxes on stage!
this modal guy is cooking rn
0
0
0
153