Former Investment Analyst. Stock trader.
New York, USA
Joined July 2026
- Tweets203
- Following78
- Followers12
- Likes327
Chip retweeted
Gina Hinojosa may not be as bad for Texas data centers as it looks.
Her plan asks for: no subsidies, union labor, no NDAs, and proof household power and water bills won't rise.
Much of that is already underway:
Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI pledged in March to bring their own power and pay for transmission
Microsoft pledged to pay full property taxes without seeking tax breaks
Nvidia is working with Vistra, NextEra and Constellation on data centers that cut power use when the grid is strained
Chip retweeted
Are Data Centers getting canceled?
With Democratic victories in Texas and Georgia investors are worried that this could bring new permitting restrictions, potentially slowing the AI buildout.
Nearly 30% of U.S. data center capacity expected to come online in 2028 is concentrated in Texas alone.
But what people are missing is that the push back on Data Centers construction is not new.
▪️ ERCOT already had a moratorium on the projects
▪️ In NY, a one-year moratorium was imposed for data centers requiring 50 MW or more.
➡️ 2027: Out of the ~40GW expected in the US in ‘27, 70% are permitted and under construction…
… Assuming typical delays, equipment availability, permitting, etc., we expect 25-30GW to come online in ‘27.
➡️ 2028: Out of the ~60GW expected in ‘28, we expect to be able to build ~35-40GW.
And this has very little to do with the election.
Bottom line is that the projects that have questionable economics, location, power source were going to come online regardless of the elections.
Separately, delaying compute supply doesn't make AI demand disappear. It makes existing capacity more valuable.
GPU rental prices were already increasing before the election.
H100 rents are up ~35% since early January, four years after launch. Newer B200 rents rose 11% in the past month alone.
If AI demand continues growing while new capacity gets pushed out, we could see GPU rental prices rise significantly further.
That has important implications for the hardware cycle:
➡️ Existing GPUs could command higher rental rates, improving returns on deployed capital.
➡️ Older GPU generations could retain their economic value longer.
➡️ Delayed construction could spread hardware spending over more years, extending the investment cycle.
The bottom line is building physical hardware is hard. Is more regulation making it harder?
Sure.
But not enough to justify 50% drop in hardware stocks.
Chip retweeted
Be Bill Ackman.
> Grow up in Chappaqua, New York.
Your dad is in real estate. Your mom is organizing commuters to fight the railroad over lousy service.
> Go gray at nine years old. Apparently, you've been stressed about the markets since elementary school.
> Dad doesn't believe in allowances, so he pays you by the hour to dig a 50-foot ditch.
> Realize pretty quickly that hourly work is for suckers. Negotiate a flat fee on the next job.
> Tell your dad as a teenager that you're going to be a millionaire by 30, worth $100 million by 40, and a billionaire by 50.
> Normal teenage stuff.
> Go to Harvard. Sell advertising for travel guides. Make $14,000 in one summer.
> Immediately blow a bunch of it on a ridiculous stereo system and learn about income taxes the hard way.
> Graduate. Work for your dad. Go back to Harvard for your MBA.
> Co-captain a rowing team that puts dollar signs on its oars.
> Get booed for it.
> Defend the dollar signs in the school newspaper.
> This guy was absolutely born for Twitter.
> Graduate in 1992. Start a hedge fund called Gotham Partners.
> Cold-call people on the Forbes 400 to raise money.
> Get married in 1994 to a woman your dad helped introduce you to by fax.
> Yes, fax.
> By 2002, your fund is winding down. Not exactly how you drew it up.
> Get into a dispute with Carl Icahn over a $4.5 million earnout from a business deal.
> Sue him. Eventually win.
> In 2004, start another fund called Pershing Square with about $54 million.
> Bet against bond insurers before the financial crisis. Make roughly $1.1 billion on the hedges.
> Buy into a bankrupt shopping mall company called General Growth Properties. By 2011, an investment that started at $60 million is worth roughly $1.6 billion.
> Not bad.
> In 2012, publicly commit to giving away most of your wealth.
> Also put on a $1 billion bet against Herbalife because you believe it's a pyramid scheme.
> Carl Icahn takes the other side.
> Naturally.
> In 2013, the two of you get into an absolutely legendary argument on live CNBC.
> Icahn calls you a "crybaby in the schoolyard."
> You're two billionaires in your 40s and 70s arguing on national television like somebody stole somebody else's lunch money.
> It's incredible television.
> Meanwhile, get involved in Canadian Pacific Railway. Push for a management overhaul. Help turn the company around.
> Exit in 2016 with roughly $2.6 billion in gains for Pershing.
> Unfortunately, also get involved in Valeant Pharmaceuticals.
> Lose about $4 billion.
> Call it a "huge mistake."
> Fair assessment.
> Your marriage breaks down. Things get rough personally. You miss your kids.
> Buy an apartment in the same building as them just to be close.
> Then start dating Neri Oxman, a brilliant MIT professor.
> She mentions that Brad Pitt is visiting her lab.
> One day she goes quiet for hours while he's there.
> You start freaking out that Brad Pitt is going to steal your girlfriend.
> Later tell Lex Fridman this was one of the scariest moments of your life.
> You've personally lost billions of dollars in the stock market, but Brad Pitt showing up at MIT is what really sends you over the edge.
> I love this story so much.
> Eventually marry her in 2019.
> Oh, and that Herbalife bet?
> Close it in 2018. Lose about a billion dollars, by your own account.
> Can't win them all.
> Then 2020 rolls around.
> COVID hits.
> Go on CNBC and tell everyone "hell is coming."
> Also tell viewers you're buying stocks.
> Your funds have put on pandemic hedges that ultimately generate $2.6 billion in proceeds against $27 million in premiums and commissions.
> One of the great trades of the financial crisis.
> Spend the pandemic buying stocks while everyone else is losing their minds.
> In 2021, donate Coupang shares then worth roughly $1.36 billion to charitable causes.
> In 2022, buy more than a billion dollars of Netflix.
> Three months later, sell it at a reported loss of over $400 million.
> Probably the most expensive Netflix subscription in human history.
> Then get into a massive public fight with Harvard over its leadership.
> After the university president resigns, write a roughly 4,000-word reaction.
> Remember that your high school yearbook voted you "most verbose."
> Absolutely nailed that one.
> In 2025, at 59 years old, decide you want to play professional tennis.
> Enter an ATP Challenger doubles tournament with Jack Sock.
> Lose 6-1, 7-5.
> Who cares. You played professional tennis at 59.
> Then decide the young men of America need dating advice.
> Go on X and recommend walking up to women and politely asking, "May I meet you?"
> The internet loses its mind.
> Double down.
> This is a man who has fought Carl Icahn, survived a multi-billion-dollar loss, stared down Brad Pitt as a romantic threat, and played professional tennis.
> You think he's worried about getting rejected on an elevator?
> In 2026, bring Pershing Square Inc. and Pershing Square USA to the New York Stock Exchange.
> And now, at 60 years old, go on CNBC and warn that the stock market has become a casino.
> Bill, my brother.
You've been digging ditches, fighting billionaires, making and losing fortunes, buying bankrupt malls, arguing with Harvard, giving away billions, worrying about Brad Pitt, and dispensing dating advice to strangers.
You've lived about seven different lives.
What an absolutely living legend! 🫡
Chip retweeted
Things in Western Europe will improve when a few hard realities are understood:
1. Europe has no God-given right to be wealthy or stable. Things can & will decline and fall apart if not maintained and adapted to new realities. Things that worked in the 1970s may be counterproductive now.
2. Energy is the basis of the economy. Without it, nothing else matters. While efficiency will be important, any economy where energy is scarce will inevitably be one that struggles
3. Not all cultures are equal. Some create prosperity, some lead to misery, some are outright evil. If Europe doesn't accept this, the cultural basis of its previous success & prosperity will be destroyed forever
4. Have kids, or die out. In the long term, few things matter this much. You can't create a European future if there are no Europeans left for it, and as the experiences of the last few decades show, non-European immigration simply makes Europe culturally less European - and doesn't magically make these people culturally European.
5. The state can not - and should not - save every individual from him/herself from him/her bad choices or ineptitudes. The idea of a nanny state which can ensure that everyone always live comfortably with minimal effort & contributions 100% of time is not feasible. The socialist promises of 30 years of wealthy retirements for everyone and Europe being able to take on endless low skill immigration were lies.
6. If a society has no way for ambitious & creative people to get rich from working hard, everyone will get poorer because of it. The hard truth is that much of our economic future depends on businesses being built & transformed by ambitious & sometimes ruthless individuals. Not everyone is able or willing to do this, and without them, there is no one to employ others and no wealth & growth to tax. But if businesses & wealth are seen solely as an evil force to be taxed into oblivion, they will go elsewhere.
7. In some cases, the best way for society to deal with a criminal is to lock them up and throw away the key. A small percentage of humans are ticking time bombs - low IQs, high aggression potential, psychopathic traits, socialised into criminal structures or a culture of lawlessness. Society can't fix them. They go in and out of court rooms and across Europe states, NGOs & social workers waste billions of hours & Euros to try & fix the unfixable. Not least because of mass immigration from dysfunctional societies, our legal systems & high trust societies built on principles such as the benefit of the doubt & correction via socialisation are overwhelmed by the amount of aggressive young men with no interest to contribute to society.
All of this is fixable. But the first step is accepting reality.
Chip retweeted
$FTAI Part Two is live!
Q3 was the best quarter in the company’s history. Aerospace margins came back to 31%. Modules hit a record. Power went from one order to more than 100 units under contract, and the customer paid $234 million before a single turbine shipped.
We read the filings line by line, and it tells a different story than the slides. The old margins were built on nearly free engines from FTAI’s own lease fleet, and that supply is almost gone. Most of the EBITDA beat came from a new definition. One customer was 37% of revenue. And the DOE rule on Chinese equipment, the biggest risk to Power, lands December 24.
The business got better. The stock got more expensive. We break down what that means and what we think FTAI is worth now.
Chip retweeted
With AI stocks down 20% to 50%, I keep getting asked whether the AI trade is dead. I was in the market during the dot-com boom in my early twenties, so let me put the two side by side.
Duration first. The dot-com boom ran 5.3 years from Netscape in 1994 to the peak in 2000. The AI boom, counted from the ChatGPT moment in late 2022, is 3.9 years old. If it merely matches the dot-com in length, there are 1.4 more years to go.
Then magnitude. The Nasdaq 100 rose about 1,000% during the dot-com boom. In the AI boom so far it is up 137%, which is roughly 13% of the way there.
Now the poster boys. Cisco was the most important stock of the dot-com era, and at the 3.9-year mark it was up around 700%. It went on to peak at 4,200% on a forward P/E of 131 times. Nvidia at the same point is up about 1,000%, and its forward P/E is 17 times. Cisco was priced for perfection. Nvidia is priced like a mature industrial.
I am not forecasting that Nvidia goes 4x from here. I am saying that on both duration and valuation, this boom is nowhere near where the last one ended. That is why I treat these pullbacks as rotations within a bull market rather than the end of one.
Learn to invest and trade at piranhaprofits.com
Chip retweeted
everyone believes intelligence has unlimited demand, not just the maxis. but what the maxis get wrong is that the intelligence anthropic and openai and frontier models are building isn’t the generic kind you’re picturing. let me explain-
the two biggest uses of compute are crudely either training or inference. the “generic intelligence” everyone says has unlimited demand comes from mainly inference, and inference needs very different power and utilization than training. inference comes in bursts, so you have to by default overbuild capacity to handle the peaks until the utilization problem is solved, which means almost by definition we are going to overbuild. but training doesn’t have that problem: the workload is steady and the demand curve is knowable. so data centers end up overbuilt right now precisely because utilization isn’t optimized yet.
so anthropic and the other frontier labs are building their future on training, which is why they are saying they can soak up all this excess capacity. but their “intelligence” isn’t workplace productivity. it’s the leap toward singularity-type outcomes: new drug discoveries, cyber capabilities, recursive self-improvement. these are true 1-of-1s. that’s why when anthropic talks about “existential threats to humanity,” it isn’t talking about job losses, not really. it’s saying they’re on the precipice of unbelievable discoveries, because they’re still TRAINING MORE MODELS.
but distillation is real and open source is real. the half-life of any frontier model is tiny because nothing protects it the way IP protects a patent. imagine a rare disease drug company that spends billions on research, and two months after a breakthrough a generic shows up. that company can never recoup its investment, unless it can discover an infinite number of rare disease drugs. THAT is the kind of intelligence anthropic is betting on. that’s what the $800bn of compute spend is betting on.
so the questions you need to ask are: 1) if intelligence is two-tiered, what does each tier cost? what’s the cost of AGI, and what’s the cost of the generic intelligence that has infinite demand? and 2) are we allocating capital with the right risk curve and decay function to find out whether the AGI payoff is real? and if it is real, could a private company ever truly own it?
that’s the cognitive dissonance at the heart of this entire paradox: everyone is gaslighting each other about what intelligence is as the unit of measurement is purposely kept vague by both the users and the suppliers. generic intelligence has unlimited demand, but it’s a certain race to the bottom. "superintelligence" is the opposite: it can't have unlimited demand by definition, but if it delivers even half of what it promises, no private company will ever get to keep it, at least not without the government stepping in to guard the moat (which is why anthropic keeps playing this angle up, without having shown any demonstration they are near this level of discovery at all). and at that point you have to ask who is going to hold the bag, and whether that was in fact, very intelligent at all.
I've become inceasingly bearish on AI / AI capex in recent months. I've worked hard to speak to VC's and founders to understand where I may be wrong. I've ultimately concluded the AI maxis believe intelligence has unlimited demand and understand very little about microeconomics or accounting.
It is eerily reminiscent of Covid when everyone clapped like seals when Fauci told lie after lie, including the double masking idiocy. We are well past that point in the AI trade. The beauty of it is, the AI maxis have chosen to lever up to the max and take out hundreds of billions of dollars of debt based largely on vibes. The die is now cast, and there is nothing anyone can do about it. Now we just sit back and watch the AI trade collapse just like Leo did.
Chip retweeted
Four weeks of wall squats reduced systolic blood pressure by nearly 10 points in adults with high blood pressure.
Researchers then compared continuing three sessions weekly with reducing to two, one, or none for another month. Systolic pressure remained roughly 7 mmHg below starting levels in the once- and twice-weekly groups. Continuing three sessions weekly produced the greatest reduction of about 14 mmHg. In those who stopped completely, blood pressure moved back toward baseline.
This fits with a previous analysis of randomized trials that identified isometric exercise, which involves holding a muscle contraction, as an effective way to lower resting blood pressure.
What interests me is the practical question - how much of an exercise benefit can we preserve when we temporarily have less time? This study suggests that reducing the frequency can retain part of the benefit for up to a month.
Chip retweeted
🚨 NOMURA: THE MARKET MAY BE MISREADING THE ENTIRE AI DATA CENTER BOOM.
Big Tech is spending hundreds of billions on AI infrastructure. Free cash flow is under pressure. Investors are worried the AI boom isn't profitable.
But Nomura argues something very different:
AI infrastructure is already highly profitable. The cash flows simply haven't caught up with the investment cycle.
Here's the breakdown 🧵👇
1/ The profitability numbers tell a different story.
According to Nomura's November 2 report:
• AWS revenue: $46B, +39% YoY
• AWS operating margin: 38.7%
• Google Cloud revenue: $30.7B, +103% YoY
• Google Cloud operating margin: 36.2%
• Azure growth: +48%
These are not the numbers of an industry struggling to generate operating profits.
2/ The real issue is timing, not necessarily profitability.
Big Tech pays for data centers, GPUs, networking and memory upfront.
Revenue and cash generation arrive over subsequent years.
That creates a classic free-cash-flow J-curve:
CAPEX accelerates → FCF deteriorates → Infrastructure generates revenue → Investment growth moderates → FCF recovers.
3/ Nomura expects operating income to surge before free cash flow.
Its illustrative forecasts suggest operating-income growth could peak around 2029, while FCF growth accelerates later.
That distinction matters enormously for investors valuing AI infrastructure businesses.
4/ Why can't hyperscalers simply stop spending?
Because AI increasingly threatens their existing businesses.
Search, advertising, shopping, software and cloud services are being reshaped by AI agents.
Companies that underinvest today risk surrendering customers, data and market share tomorrow.
5/ Demand may not be the primary constraint. Capacity is.
Nomura cites enormous contracted cloud backlogs:
• Google Cloud: $560.8B
• AWS: ~$496B
• Microsoft commercial RPO: $677B
These figures have different reporting definitions, but collectively illustrate substantial contracted demand.
6/ Big Tech has financing alternatives to cutting CAPEX.
Nomura identifies four paths:
① Accelerate investment
② Freeze CAPEX
③ Use joint ventures and operating leases
④ Monetize contracted future cash flows
Its preferred expectation is continued investment supported by increasingly sophisticated financing structures.
7/ Oracle offers a real-world example.
Nomura highlights Oracle's combination of long-term leases, project financing, customer prepayments and customer-provided hardware.
The broader message: hyperscalers can restructure who finances infrastructure without necessarily reducing how much infrastructure gets built.
8/ This distinction is critical for semiconductors and memory.
Negative hyperscaler FCF does not automatically imply collapsing demand for GPUs, DRAM, HBM or other data-center components.
If spending continues through financing structures, semiconductor demand can remain resilient even while hyperscalers optimize their balance sheets.
9/ What could invalidate the thesis?
A genuine compute oversupply.
Watch:
• Cloud revenue growth
• GPU rental pricing
• Contracted backlog growth
• Data-center utilization
• Memory supply commitments
• Hyperscaler investment guidance
These are more useful indicators than negative FCF alone.
10/ The investment takeaway:
The key question is not simply:
"Why is Big Tech burning so much cash?"
It is:
"Will the infrastructure being built today generate enough durable cash flow to justify its cost of capital?"
Nomura is optimistic.
But the answer ultimately depends on pricing, utilization, investment discipline and the durability of AI demand.
The AI infrastructure cycle may be a cash-flow timing problem rather than a profitability crisis.
$NVDA $MU $GOOGL $MSFT $AMZN $META $ORCL
Chip retweeted
I feel like I just read a Ferg post. 1000% agree with the thesis. Momentum right now is lower but PPLT and PALL are definitely great long term portfolio adds.
Chip retweeted
2nd post is about commodities and what is "cheap" right now. As a general matter of fact, a commodity for me is cheap, when 30% of producers cannot produce at a profit. It is very cheap, if that number is 50% or more. You should then cross-check if that commodity might be in a structural or cyclical demand downtrend. Why? Well, the price might go down for a reason. E.g. diamonds: structural downtrend, (nearly) no-one needs (natural) diamonds anymore so diamond mines are closing and for a reason. These companies represent what people call a "value trap". Cyclical downtrend: good example is lithium: a lot of unwarranted optimism about EV adoption, lithium price skyrocketing, but a LOT of lithium out there so many projects quickly in development, more supply, EV adoption a lot slower than expected: cyclical oversupply, lithium miners are down and with a reason. They then go down even more, overreacting to the negativ and then will go up again some day with supply being lower than demand. So thats the exceptions you should always crosscheck! Now if you look for gold, oil etc. you can see that ALL producers now are producing with positive FCF. Is oil or gold therefore cheap? Not in the short-term. Oil is cheap in the longer-term though as the world needs an oil price that incentivizes more drilling, but that is longer-term and short-term everything can happen. As for all miners, costs are up a LOT due to higher oil/diesel prices and higher wage costs. Silver: Times where the average silver miner produced at a FCF break even of 25 USD are long gone. But I guess at 49 USD nearly all silver producers are still producing profitably, so is silver cheap short-term: not really. Longer-term the world needs more silver, so it needs a higher price to incentivize silver exploration and mine development. Now for the more interesting commodities: platinum and palladium. I posted the cost curve for 4E production from Sibanyes website. Its a little bit noisy as 4E includes palladium, platinum etc. and it also only includes SA producers and Sibanyes US production, but e.g. no production from Russia. 43% of palladium production comes from Russia and around 13% of platinum production. One thing can be said though: with all SA producers having costs of around 1450 USD/oz of "4E" right now, no-one of them makes a profit at current PGM prices. I tried to calculate the FCF break even of Russias production, which is very difficult given that e.g. Norilsk Nickel also produces nickel, zinc etc besides PGM. But given the fact that especially nickel is down as well, break even for Russia´s palladium should sit at around 760 USD/oz. So they are still profit making = around 45% of palladium production and lets say 20% of platinum production currently is profitable. Now SAs PGM producers all are more or less net debt free and have very stable balance sheets. Means they can withstand a long time of losses before being forced to shut a mine and reduce supply. Plus car demand is slowing due to "the economy" = higher interest rates etc. ((PGM are mostly used in catalytic converters)). But one can say that with 85% of the worlds platinum industry and 55% of the worlds palladium industry currently producing at a loss, PGM are definitely cheap and their price is not sustainable here for the longer term. 3rd post following.
Sick guy who spams you through every channel by creating multiple avatars of himself even if you block he will spam. And his “levels” are useless.
For today only, I'm giving followers who message me on WhatsApp my U.S. stock trend-quality notes and chart levels. See what makes me question a fast price move. Tap below to get today's update 👇 watchlistandlevels.com/p/fp4…
Chip retweeted
Our deepest research piece… and arguably the most consequential is now published. $ORCL
We take a look at Oracle, the company at the center of the largest infrastructure buildout in human history, its financials, and its potential as an investment:
kryptonresearch.substack.com…
Chip retweeted
you're seeing a consistent pattern:
- Fortinet: people bought bc of AI cybersecurity 'emergency'. turns out 4q decelerating, management price sensitive. stock -25%
- ServiceNow: AI SAAS transformation requires forward deployed engineers, turns out 4q is decelerating - stock -25%
- Google: spent entire time on TPUs and AI, going FCF negative while its core business (ads) is - you guessed it, decelerating in 4Q. stock -13%
this was after youtube was supposed to be killing Netflix (stock -44% over past year)
- Robinhood: big agentic push, Monthly active users sequentially decelerating.
meanwhile the consumer sector (big ad buyer) is in beast mode. Cheesecake factory 15% comp, Chili's 10%. Norwegian cruise lifts full year guide - passengers up 22% year over year (983k vs 900k expected).
which begs the question: if you have RSI, and AI is the future why are you underperforming fried mac n cheese sales or cruise billings in the middle of a war. like what are you actually investing against
so company after company is going full tilt into AI, and not really getting a return - just as everyone is
1] locking in to multi year memory contracts (which is part of the price sensitivity fortinet noted)
2] going FCF negative
3] ready to absorb $2-4T of public equity issuance from the labs
meanwhile market rejoices when Oracle *cuts back* investment guidance in data centers.
Labs on fund raising roadshows are saying the new demand story is that software and hardware are insecure and need to be completely refactored for AI era. Which is just an expense, with no revenue acceleration - which brings to the macro question of 'is productivity going to accelerate'
right now sub 2.5% - needs to be above 4% sustainably for debt dynamics to work out. if everyone has to rewrite their software to address cyber threat environment, and banks get hacked constantly (which is apparently the bull case for labs?) - you enter unustainable fiscal situation and decelerating growth which was buttressed by AI infra investment which (per Oracle) is likely to decel
even at macro valuation level, China is far ahead of Europe in AI (non controversial). Hang Seng Tech index trades at 10.5x forward PE while French stock market (CAC Index) trades at 14x while Chinese 5 year yield is 1.3% and France is 4.3%
so there's a completely believable world in stocks where you get a lot more inference, and a lot more AI, and multiples decline or costs accelerate. which seems like Blasphemy in SF but if you just listen to public companies seems closer to a base case
Chip retweeted
Are Investors Finally Waking Up To "The Great Rare Earth Disconnect"? zerohedge.com/markets/are-in…
Chip retweeted
A lot of ink has been spilled on GPU financing lately. I spent over a decade trading ABS on Wall St, and structured over $2B+ of debt financings as an operator (including at @tryramp). Figured I'd try to make sense of this market, so I went back to a framework I knew well: aircraft ABS and aircraft engine ABS.
My take: a GPU neocloud runs on airline economics, and the debt behind it looks a lot like aircraft engine ABS.
The parallels:
• Perishable inventory. An empty seat after the door closes is gone forever. So is an idle hour on an energized H100 rack.
• 85%+ of cost is fixed (servers, 10-year rack leases, 10-14% debt service). One more hour of compute costs pennies of power.
• Utilization is everything. Airlines break even around 80% load factor, a levered GPU cluster around 65-75%. The returns live in those last 25 percentage points.
• NVIDIA is Boeing and Airbus rolled into one. Powered shells and grid interconnects are the new airport slots.
• Wall St copied the aircraft financing playbook: take-or-pay contract plus servers in a bankruptcy-remote SPV, 70-80% LTV.
Part 1 of my blog post goes into a lot more detail:
alexyangsong.com/blog/alumin…
Chip retweeted
A very sharp friend of mine sent me a take arguing that Cerebras chips could depreciate extremely quickly. I thought I’d share it here.
Cerebras’ wafer-scale engine (WSE) and routing architecture do seem fully capable of solving the yield problem, as the company claims. The uncertainty is around how long the wafer will actually last in operation.
Cerebras’ design requires a large number of redundant interconnects to accommodate manufacturing defects. But all that BEOL wiring significantly increases the effective area susceptible to TDDB (time-dependent dielectric breakdown), potentially creating much greater long-term reliability risks. TDDB can cause shorts, which are more dangerous than opens in this design.
Dielectric thickness determines the probability of TDDB, and variations in that thickness can be spatially correlated across a wafer. That makes it highly likely that many interconnect failures would cluster in a particular region. It is unclear how Cerebras detects these shorts and avoids using the failed interconnects.
Ultimately, the argument is that you have a design that is expensive to begin with and also ages extremely quickly.
Chip retweeted
Commented on $ASTS - AST SpaceMobile: Story Collapses Before Even Starting. seekingalpha.com/article/495…
Chip retweeted
Many of you don't understand how serious this is, so let me break it down.
In computer science there is a famous kind of puzzle called 3-SAT. You are given a long list of yes-or-no switches and a set of rules about how they must be arranged, and you have to find a combination that satisfies every rule. The obvious way to solve it is to try every combination, but each new switch doubles the number of possibilities, so a puzzle with a few hundred switches would take longer than the age of the universe to brute force.
For decades, the smartest people in the field found clever tricks that made this a little faster, but every trick still hit the same wall: the time required kept exploding as the puzzle grew.
About 25 years ago, two computer scientists, Russell Impagliazzo and Ramamohan Paturi, working with Francis Zane, turned that wall into a formal idea.
They proposed the Exponential Time Hypothesis, which says this explosion is not a failure of human cleverness but a law of nature. No algorithm, however brilliant, will ever solve 3-SAT without the time blowing up exponentially. Nobody could prove it, but almost everyone believed it, and it became one of the foundations of modern computer science.
Researchers used it as a measuring stick. If you could show that solving some other problem quickly would also crack 3-SAT quickly, you had shown that problem was hopelessly hard too. Hundreds of papers, across scheduling, graphs, logic and much more, rest on this one assumption.
Now an internal AI model at OpenAI claims to have broken it. It says it found an algorithm that solves these puzzles meaningfully faster than the hypothesis allows, and it has checked the proof in a computer verification system called Lean.
And here is the remarkable part: the very people who proposed the hypothesis, Impagliazzo and Paturi, along with MIT's Ryan Williams, are now leading the public review of whether a machine has disproved their own idea.
So what happens next? If the disproof holds, a large part of theoretical computer science has to be revisited.
Every result that said "this problem cannot be solved faster, assuming ETH" loses its foundation, and researchers will have to work out which of those walls were real and which were simply where we stopped looking. It does not mean everything becomes easy. The new algorithm is still slow in absolute terms and this is not the famous P versus NP problem being solved.
But it means our map of what is possible was wrong in a place we were very confident about, and it was an AI that redrew it. If the disproof fails, the hypothesis survives and the field breathes out, but we will still have watched an AI produce serious, checkable work on one of the deepest questions in the discipline, closely enough that the founders of the idea had to take it seriously. Either way, something has shifted.
