there's a lot of FUD on the timeline this week, so i went back to my favorite story about how fear actually spreads.
three schoolgirls on a train once started a run on a perfectly healthy bank.
Japan, December 1973.
one of the girls had just landed a job at Toyokawa Shinkin, a small local bank. her friends teased her that banks were "dangerous" (they meant robbers).
it was a joke. she didn't take it as one.
that night she asked a relative if it was true.
The relative asked someone else, and the question passed from family to family, through a beauty salon, picking up weight as it went.
Then a man borrowed a shop phone and told someone to withdraw 1.2 million yen for a routine payment. a listener overheard and assumed he knew something.
now the rumor had a witness.
a taxi driver later traced how the story changed in a single day:
- afternoon: the bank "might be in danger"
- 2:30: "it is in danger"
- 4:30: "it will go into bankruptcy"
- night: "its shutter will never be lifted tomorrow"
nobody added a single fact, each retelling just dropped a hedge.
by the end, more than a billion yen had walked out the door. The Bank of Japan rushed in cash and stacked it a meter high in front of the vault so people could see it with their own eyes. Police went looking for the crime behind the rumor and found nothing, the bank had done nothing wrong.
so why does the scary version always win?
because a loss feels about twice as strong as an equal gain. Psychologists have measured it for decades. Standing in line costs you an afternoon, but being the one who stayed could cost your savings. with that trade in front of you, running is the rational move, even when the bank is fine.
which brings me to this week.
the timeline is busy with stranded compute and debt stories. Those are fair questions, and some builders will deserve them, but the rumor doesn't check who, it hits everyone standing closest.
so do what the Bank of Japan did, skip the gossip and go look at the vault.
is the power actually connected?
are the customers actually signed?
is the money already there?
those answers don't change because a post went viral. They change slowly, in substations and contracts and concrete.
and that work didn't stop this week. If the cranes are still up, the racks are still getting installed and the demand that started all of this is still there.
a run doesn't make a bank unsound, it just tests whether it was.
Toyokawa survived, the ones building on solid ground usually do.
Show more
Susan Li, Meta’s cfo, pulled Zuckerberg aside at an exec offsite in hawaii late last year. she wanted to make sure he understood that if this model didn't work, it would be very, very bad in financial terms. Fun fact is nobody had seen the model yet
and the bet was already poured in concrete. Prometheus in ohio is past a gigawatt and hyperion in louisiana scales to 5, and both were planned in 2023 or 2024, before the new lab even existed. @alexandr_wang called it the opposite of lean startup, you double down before you know anything
zuckerberg's answer to the "very bad" scenario is the most interesting risk take i've heard from a hyperscaler this year.
For it not to work out, "it's not just that we have to fail. AI overall has to fail." if meta's own model flopped, plenty of other people would want that compute and pay more than it costs to run, and meta still had other internal uses for it anyway
so the hedge was never muse. it was the GPUs.
Meta committed to sites with power before the product existed, knowing the floor was the open market for compute. you don't call idle capacity a backstop unless you expect a line of buyers, which is a pretty big tell about where they think rental prices are going
Show more
the TPU is still the star of google's ai story, but the chip sitting next to it might decide how fast your answers come back.
$MRVL investor day yesterday spelled out its google deal: custom chips that plug into the TPU world, inference accelerators included, worth up to $120B through 2033 if Google hits its purchase targets.
a second inference only chip reportedly paired two to one with tpus (treat the ratio as rumor)
it makes sense once you split a reply into its two jobs:
1. reading your prompt is a math job, and a TPU or GPU eats it for breakfast
2. Writing the answer word by word is a memory job, because for every token the chip has to pull the model and your whole conversation back out of memory.
Etched's CEO @UbertiGavin said it plainly on invest like the best: decode is all a memory game. on blackwell, he said, one hop between gpus can take about 4,000 nanoseconds, which is why eight gpus don't make you eight times faster.
So the hardware is splitting along that line.
One chip for thinking about the prompt, another built to talk back fast. That puts the pressure on memory bandwidth across the whole cluster and on the wires between chips.
No surprise that interconnect is about $37.5B of the growth $MRVL laid out
Show more
Google's deal is one big ass contract.
The firm power map around it is a bit bigger:
1. Constellation $CEG - largest US nuclear fleet; Google uprates plus Amazon on Calvert Cliffs (~190 MW) last week
2. Talen $TLN - Susquehanna already feeding AWS (up to 1.9 GW through 2042), roughly 4 GW of DC sites in the pipeline
3. Vistra $VST - rode the same bid on ERCOT/PJM firm capacity
4. Then the "you still can't interconnect" layer: Hubbell transformers (lead times still talked about in years), Quanta / EMCOR for the MEP that wires the hall
5. Bloom $BE if the queue won't move - onsite fuel cells, Fremont racing toward 2 GW annual by year end
Show more
had a bit of a "wow" moment, super humbling
spent some time looking at who’s reading and using AI Bottlenecks, and here are some of the logos
got subscribers from investment firms, venture funds, semi companies, AI labs, and teams building the infra behind all of this.
back in May, I couldn’t sleep, so at 2 am. I started AI Bottlenecks as a pet project. I wanted a place to work through a question I kept coming back to: what could actually slow down the AI buildout?
Since then, website got hundreds of thousands of visitors, the posts made millions of impressions.
People have started following, reading the newsletter, using the app, and getting in touch. I’ve had conversations I never would have expected when I first sat down to build it.
It’s humbling to know that these people find this useful enough to keep reading.
These are individual subscribers, of course, not company endorsements. But to everyone who has read, shared, used the app, or sent me a note: thank you 🫶
It still feels like a pet project to me, and I’m grateful you’re here for it
Much more to come...
Show more
This is unique and a fun, short project. I spent some time looking through my subscribers emails and found this.
One thing that is common is how many people from investment & financial firms are tracking us.
Very cool...
Show more
ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in our European datacenters -- our cluster in Bruyères-le-Châtel built using our Series B fundraise.
We are investing heavily in infrastructure, with our Series C and D clusters coming online soon. This additional compute will let us train longer, pursue more ambitious post-training, and iterate faster.
We expect large and rapid improvements in the weeks and months ahead.
2/n
Show more
damn i think i created a monster
- long only, no shorts, no leverage (sorry leopold)
- $10k clean slate, sep 12
- tracks 223 names (ai, energy, robotics)
- research first, trade after, limited risk taking
- one pass, 223 decisions, no cherrypicking
- deterministic allocator
- opportunity rank 1->223
- 25 max positions, 6 exits a session
- minimax M3 for trading decision
- research with minimax + GLM + kimi
- +5.33% since inception
- exposure 74%, 36 names, $2.7k cash
Show more
10 days ago, i gave my LLM $10k to trade with and a ton of research tools as an experiment
the system is built on top of what i built with and uses:
- X signals + X websearch(manual selection of top voices + automated search)
- @perplexity_ai research
- @tavilyai
- Company filings
- Earnings, analyst targets etc
- News
- Technical and catalysts
- private reports
I limited the scope to 223 stocks in AI infra, energy and robotics. But I didn’t ask it for "stock picks", I asked it to run a portfolio.
Every session, MiniMax M3, running through @nebiustf, must decide what to buy, hold, add, trim or avoid.
For every stock, it produces:
→ A target weight
→ 1, 5 and 20 session forecasts
→ A stop loss
→ A profit taking ladder
→ A time stop
→ A rationale, counterargument and invalidation trigger
Then a deterministic risk engine gets the final word.
It checks concentration, correlations, basket exposure and position size.
Every call is timestamped, every trade fills at the next observed close, and costs are included.
Nothing gets rewritten after the fact.
The most interesting trade so far?
$AMD
The bot started building the position on September 15.
A few sessions later, AMD had jumped 9.9% in one day on 2.3x normal volume. The position was up 22.6%.
The thesis was still improving.
But RSI had reached 72.9. The analyst consensus target had effectively caught up with the price. And the AI basket was already at its exposure limit.
The model said hold, but the risk engine trimmed the position anyway.
That is exactly the behavior I wanted! It liked the company but refused to let conviction become concentration.
The portfolio it built includes:
Compute: Micron, Nvidia, Broadcom, TSMC, Monolithic Power
Power: Siemens Energy, BWXT, Caterpillar, First Solar, Mitsubishi Heavy, ABB, Cameco
Robotics: HIWIN, Novanta, Tesla
Early results, from a clean $10,000 start:
Portfolio value: $10,471.69
Return: +4.72%
Ahead of its benchmark: +4.31 points
Executed trades: 114
Cash: 10%
Very early. Most positions have only four or five observed sessions.
And the first version already taught me something:
The bot found too many “good” ideas.
It opened 63 positions, often in tiny amounts.
So I kept the same account and the full history, but changed the policy.
It is now concentrating toward 25 positions, with larger minimum entries and one in, one out replacements.
The experiment isn’t really wether or not a LLM can predict stocks, it’s more "can a model turn a mountain of messy info into repeatable, constrained and fully auditable decisions?"
Now I can finally measure it
will share more as the experiemnt continues...
(fictive money, not financial advice, just a fun experiment)
Show more
Google just signed a 20 year nuclear PPA with $CEG
3,590 MW, 890 MW of new uprates at 11 PJM units, >$4.3B of investment.
CEG +12.9%.
The message is clear: hyperscalers will pay premium for power that never stops.
Existing nuclear just got repriced from boring utility to scarcity asset.
$VST +9.5% on the $4.2B DOE loan for 433 MW of uprates.
Show more
the chinese open weight frontier has been running without a serious western peer for a while, but today reflection dropped beam!
underneath there's a super solid team:
- ex-deepmind founders (@MishaLaskin / @real_ioannis)
- $4.6b raised,
- $25b mark, nvidia in the round.
- 23.8t pretrain on 6,144 GB300s
- then 10.5k gb300s for four weeks of rl
- 100m+ rollouts
- locked spacex colossus GB300 capacity
- locked $1b $NBIS compute deal (we announced this few months ago, letsgoooo)
- 501b total, 23b active, sparse MOE, trained from scratch, coding and agents first.
On their numbers it basically ties glm-5.2 on the agentic coding rows that matter for a workhorse:
- deepswe 44.4 vs 44
- terminal-bench 80 vs 81
but all this while using roughly 3 to 4x less inference compute for similar reasoning scores, quite cool!
it still trails glm-5.3 and deepseek v4.1 flash on those same benches tho (flash is a different animal anyway: tiny active count, multimodal). And yes kimi k3 stays ahead on raw capability.
So yes the western push is real, but part of product is efficiency (at least on this one)
closed frontier (opus / astra lvl) still owns the top of the capability curve but Beam isn’t pretending otherwise.
i guess it's fair if you compare open to open, and having a western model you can own that sits in glm 5.2’s neighborhood at a fraction of the reasoning compute, is quite impressive
as Laskin said: closed models are renting an apartment, an ownership market is forming. This is the first well-funded western open workhorse built for that
Show more
We are introducing Beam today, a 500b open model that excels in coding, agentic, and scientific workloads.
Beam is very token-efficient, pre-trained from scratch, and scaled with the largest RL run documented openly we are aware of.
Getting here was challenging, rewarding, and personal.
I immigrated to the USA during childhood when my parents got jobs as scientists at PNNL, a national lab located in a rural town in Washington state.
Growing up, I developed an interest in physics because it was grounded in simple principles, was foundational, and had long-term impact.
Physics was the foundational science of the 20th century. The first general computer (ENIAC), the transistor, GPS, and much of modern technology had its roots in physics.
In 2016, during my graduate studies, I noticed the birth of a new foundational science - Artificial Intelligence.
AlphaGo, the superintelligent Go agent developed by DeepMind, had come out and made me feel what was about to come.
A neural network, trained to imitate human players, and then improve itself through trial and error, mastered the most intellectually challenging board game in the world.
What surprised me about AlphaGo is how it was conceptually simple, general, and scalable. The system could in principle improve itself forever, it was just a function of compute.
This is when I became “AGI pilled” as they say.
I switched from physics to AI because I thought this would become the root node science of the 21st century, much like physics was the root node science of the 20th century.
I believe that working on foundational science is not enough on its own. Foundational science must also remain open and widely accessible.
We remember Newton for his contributions to physics, but he was also an alchemist.
Alchemy was done in secret, because the artificial production of gold was deemed lucrative but also dangerous with the potential do destroy entire economies. Newton even wrote his alchemy notebooks in cypher.
On the other hand, Principia was published in the Philosophical Transactions of the Royal Society, the knowledge was made widely available, passed down for generations in what we now call Newton’s laws.
The computer, transistor, GPS were possible to build because physics as a science was open. Openness creates an ecosystem of innovation.
The notion of openness, which is a core tenet of scientific discovery, has also become a guiding principle for how foundational technology gets built and distributed.
The internet was built on open protocols. The most widely used operating systems are open. The field of cybersecurity exists because strong encryption protocols were made open in the 1990s.
Open also means safer. The best defence is a distributed one. Commonly used open source software is robust and battle-tested because it is inspected and patched by a community of developers.
AI is different but also similar to software in many ways, though with a much larger surface area of vulnerabilities. It is impossible for a small group of safety researchers working in secrecy to remediate all of the potential issues regardless of their intentions. It is like entrusting a handful of white blood cells to protect your body.
AI is both foundational as a science and a technology, and the safest way that also ensures the benefits of AI are distributed evenly is by making it open.
I’m excited for what developers and scientists build and learn from this model, and for Beam to play a small part in ensuring that AI is made widely accessible, safe, and beneficial to all.
Show more
if you're a quant, PE analyst, family office, or anything along those lines, please reach out! 📩
looking for beta testers users for something i’m cooking
this account is a goldmine
Deutsche Bank just initiated Formfactor at Buy with a $200 target.
They’re saying FormFactor is becoming a real second source for Nvidia GPU probe cards at TSMC, with room to take 20% share by 2028 and more than $200M of sales from that alone.
Probe cards are the boring thing on every HBM wafer and every GPU package. When Micron locks most of FY27 HBM, the test queue gets tighter.
-> last quarter probe cards did $209.7M inside $258.2M of revenue.
-> Q3 guide is $270M.
$FORM $MU $NVDA
Show more
today we announced the acquisition of Inferize.
i guess headline is the obvious part. Behind it is a bigger bet on what @nebiustf becomes as a production inference platform.
here's the problem Inferize was built for.
-> When you launch a model, spin up for a demand spike, or reload weights mid run (think reinforcement learning), the GPUs are assigned but not yet useful. That warmup window is an idle tax. Platforms hold spare capacity just to stay ready for the slo. You pay for metal that isn't shipping tokens yet.
@InferizeAI's technology cuts that tax. Faster time from "we need more capacity" to "it's actually serving." utilization tracks real demand tighter. Token economics get better because more of the rack does useful work.
The team joins Token Factory with deep gpu systems chops. Running inference well is the whole system responding when demand changes, not just fast GPUs and a tuned model.
this also fits the stack we've been assembling on purpose.
1. Eigen brought optimization at the model, kernel, and system layers
2. Clarifai's core team and licensed tech added system-level inference and compute orchestration
3. Inferize is the layer that attacks readiness and elasticity.
3 moves, one job: serve more customer demand from every GPU, and make TF the place open-weight / production inference scales without burning money on idle readiness.
that's the potential i'm excited about (not a logo on a press release). A Token Factory that gets more elastic, more utilizable, and harder to outrun on the economics of being ready.
Show more
an agent can write code in seconds. But installing dependencies, running tests and waiting for results can take MUCH longer
When you’re training agents, all that waiting is important. Faster text generation only speeds up one part of the process.
My colleagues Di and @Mhr1036 at @nebiusai worked on the rest: how tasks get routed, how tools run, and how completed work becomes a batch the model can learn from.
One idea I love: use groups of completed attempts as they become ready, so unrelated slow tasks don’t hold everything up
Across the changes, they reduced the time to collect a training batch from roughly 10 minutes to just over 3 in a controlled Terminal Bench based test, with model weights held fixed.
That’s 66.9% less time gathering the next batch
For me, this is an interesting part of building better AI: the engineering around the model determines how much of its speed becomes useful progress.
Full breakdown below
Show more
Why don’t we throw all the world’s compute at the one problem we all agree needs solving?
Because compute can’t optimise what we haven’t learned to measure.
What can be measured can scale, what scales begins to shape the world.
Show more
10 days ago, i gave my LLM $10k to trade with and a ton of research tools as an experiment
the system is built on top of what i built with and uses:
- X signals + X websearch(manual selection of top voices + automated search)
- @perplexity_ai research
- @tavilyai
- Company filings
- Earnings, analyst targets etc
- News
- Technical and catalysts
- private reports
I limited the scope to 223 stocks in AI infra, energy and robotics. But I didn’t ask it for "stock picks", I asked it to run a portfolio.
Every session, MiniMax M3, running through @nebiustf, must decide what to buy, hold, add, trim or avoid.
For every stock, it produces:
→ A target weight
→ 1, 5 and 20 session forecasts
→ A stop loss
→ A profit taking ladder
→ A time stop
→ A rationale, counterargument and invalidation trigger
Then a deterministic risk engine gets the final word.
It checks concentration, correlations, basket exposure and position size.
Every call is timestamped, every trade fills at the next observed close, and costs are included.
Nothing gets rewritten after the fact.
The most interesting trade so far?
$AMD
The bot started building the position on September 15.
A few sessions later, AMD had jumped 9.9% in one day on 2.3x normal volume. The position was up 22.6%.
The thesis was still improving.
But RSI had reached 72.9. The analyst consensus target had effectively caught up with the price. And the AI basket was already at its exposure limit.
The model said hold, but the risk engine trimmed the position anyway.
That is exactly the behavior I wanted! It liked the company but refused to let conviction become concentration.
The portfolio it built includes:
Compute: Micron, Nvidia, Broadcom, TSMC, Monolithic Power
Power: Siemens Energy, BWXT, Caterpillar, First Solar, Mitsubishi Heavy, ABB, Cameco
Robotics: HIWIN, Novanta, Tesla
Early results, from a clean $10,000 start:
Portfolio value: $10,471.69
Return: +4.72%
Ahead of its benchmark: +4.31 points
Executed trades: 114
Cash: 10%
Very early. Most positions have only four or five observed sessions.
And the first version already taught me something:
The bot found too many “good” ideas.
It opened 63 positions, often in tiny amounts.
So I kept the same account and the full history, but changed the policy.
It is now concentrating toward 25 positions, with larger minimum entries and one in, one out replacements.
The experiment isn’t really wether or not a LLM can predict stocks, it’s more "can a model turn a mountain of messy info into repeatable, constrained and fully auditable decisions?"
Now I can finally measure it
will share more as the experiemnt continues...
(fictive money, not financial advice, just a fun experiment)
Show more
Coming from the goat, means a lot 🐐
