Web 1.0 VC & BTC Hackathon ‘13. Also neuroscience, law, games, sports tech. List maintainer; I may not follow you, but you’re probably on one of my lists.
$NVDA concerns along with suspicions raised by Jensen’s number of direct reports, lack of succession plan, handshake culture, nepotistic hiring, and DELAYED 10Q.
@PTrubey@MNa409@charles12388 The theory is that US customers have the demand, but need to gate GPU acquisition to not more fully out pace electric power provision than they already do (you’ve no doubt heard about chips sitting in warehouses). SE Asia provided a sink for the excess supply.
Go read this post 👇 from March 22, 2026 — eight months ago, replying to @nntaleb — because everything that just blew up between Nvidia and TSMC this week was already written down right there.
The five-second epistemology of what just happened to Nvidia isn't spy novels, Washington lapel pins, or palace coups. Make the whole picture about TSMC and it takes five seconds to see:
What just happened is a textbook bullwhip.
Fact-check the Ops math yourself (Jay Forrester at MIT in 1958, Hau Lee at Stanford in 1997): a bullwhip happens when small distortions at one end of a network amplify 200% to 500% upstream — driven by shortage gaming, order batching, tariff/price games, and zero buffer slack.
That is specifically why the kitchen has always said they shouldn't even be called "supply chains." A chain is rigid iron — pull it an inch, the other end moves an inch. Real-world manufacturing networks are SUPPLY SPRINGS. Stretch a spring with thirty years of stripped-out buffers, load it with fake credit, and pull it past its yield point — what happens?
A spring breaks. And that crack at the tip is your bullwhip.
Look at this entire mess from TSMC's seat, because TSMC is the physical anchor bolted to the floor at the far end of the spring:
• Remember, Tim Cook and Apple built modern TSMC by paying cash like clockwork (~35 DSO) without ever yanking the spring. Nvidia did the exact opposite: every U.S. neocloud and hyperscaler double-ordered GPUs on $680B of borrowed money BEFORE they had permits or grid power (the exact energy bottleneck from March — unpermitted gas turbines in Memphis, generator fines in New Jersey, Oracle declaring force majeure over power delays).
• Nvidia took that inflated signal, stretched the upstream spring to $390B in TSMC and HBM supply commitments, extended $81.03B in customer IOUs (64 DSO), and shipped 600,000 GPUs (~$16B) into a broken Washington tariff detour where they froze unpaid in Hong Kong port.
• No slack. No cash buffer ($29.09B pure cash left after buybacks vs. $21.07B in unpaid supplier bills, up 115% YTD).
So the spring snapped right at TSMC's door.
That is why TSMC is in Santa Clara HQ and walking U.S. data-center sites demanding cash upfront. That is why Nvidia's ten outside directors — who aren't evil masterminds, just scared HOA-committee types built to form task forces and hire lawyers whenever somebody yells "China!" — got completely blindsided.
And that is why Jensen resigned five days before Q3 earnings while Nvidia hid behind an uncertified 8-K with no SOX Form 10-Q. You can't criminally certify revenue on a snapped supply spring.
None of this is complicated. All of this was written to you on March 22. Bullwhip.
@PTrubey@MNa409@charles12388 The theory is that US customers have the demand, but need to gate GPU acquisition to not more fully out pace electric power provision than they already do (you’ve no doubt heard about chips sitting in warehouses). SE Asia provided a sink for the excess supply.
In my ongoing quest to find the limits of Opus 5.5, I directed it to audit my old scientific field, theoretical astrophysics. We indexed about 150,000 papers, audited 50,000 for issues, and did progressively more thorough checks on thousands more. We found nearly 1000 major errors that got cited many times, most are still uncorrected. This survey enabled me to estimate the efficacy of peer review for correcting errors (not very good) and to identify dozens of open questions that anyone with spare Claude Max tokens could answer today (.md context files included).
Bottom line: AI can speed up quantitative science by >1000x with no further model improvement required. But I'm seeing no evidence of this being pushed within universities, only academics complaining about labs dropping math proofs.
Among my fellow reformed academic high agency founders, however, several are speed running entire disciplines.
This survey has implications for how science will be run in the very near future. Hint: Academic publishing is dead. It was already a very stupid way to record and promulgate knowledge, and had been so for decades, even before all that publicly funded knowledge was locked behind paywall cartels. Now AIs can generate thousands of novel, correct papers per day, it is no longer meaningful for humans to consider a research paper a "unit of science" or to even bother much with the 99.9% of science which, up to this point, was necessary drudgery. Time to discover something like General Relativity or Evolution every week. LFG!!
Enjoy: https://t.co/BhRxKVvbim
A lot of people have real trouble thinking about the aggregate behavior of groups. They can only think about individuals.
So they think about Elon Musk as an individual. Maybe they think that he must be retarded, because they need that for emotional reasons. Or maybe they think he is a talented rocket designer.
But even if they realize he understands rockets, they can only think about him as an individual doing tasks. Not about his impact on what the whole company does.
If Elon Musk is just a talented rocket designer, then he can spend all day, every day, improving rocket designs.
But he is just one of 7000 engineers. So even if did a perfect job, he would only improve progress by about .014%
So, when a midwit looks at SpaceX, he sees 7000 engineers, and says to himself "Elon Musk cannot possibly be designing the whole damn rocket himself. These people are giving him too much credit."
This would be absolutely true if Musk were just a some random SpaceX engineer. But he isn't. He's the absolute dictator of the whole damn company.
He decides, with absolute and unquestionable authority, which behaviors are punished and which are rewarded.
Why is this important?
When I worked at Boeing, on various classified projects, I naively came in with the assumption that I was a software engineer, employed to design and build functional software as quickly and cheaply as possible.
Because of this, I focused on functionality and efficiency, for example suggesting changes that would obviate the need for entire projects and software components.
And I very rapidly became persona non grata.
At first, this caused me to adopt an air of aggrieved innocence. I was just trying to do my job well, and I was being punished for it.
Then I gradually came to realize that I hadn't properly understood what my job was, or what the actual mission of Boeing Defense, Space, and Security was.
Boeing DSS was not an aerospace company. Boeing DSS was a fraud company.
Its goal was twofold. First, to maximize transfer of public tax dollars to stockholder pockets. Second, to maintain the appearance of being an aerospace company, to prevent the fraud from being exposed.
Which meant that my real job was twofold. First, to generate billable hours. Second, to look busy.
This is why Space Launch System sucks. Why it is always late when it doesn't fail outright. Why it is at 300% of its target budget. Why it leaves American astronauts stranded and needing a lift from SpaceX.
That's because, for Boeing, building rockets is a side quest. Something they have to do to look busy. Yes, a rocket does occasionally fly, but that's just happening so programs don't get shut down.
For Boeing, coming in early and under budget would actually be failure. The system is functioning exactly as designed.
The cost, not the rocket, is the goal.
Musk and SpaceX don't work like that.
Unlike Boeing executives, Musk is not a useless grifter asshole. He is an entirely different kind of asshole... a rocket obsessed sperg.
So, at SpaceX, engineers do not get rewarded for generating billable hours. They get rewarded for making rockets fly. And for all the subtasks of this, like figuring out how to make rockets lighter, figuring out how to make engines reignite better, figuring out how to make rocket exhaust not destroy launch pads, etc.
And this attracts engineers who care about rockets. You've seen them on broadcasts, crying on open microphones because they dedicated their lives to something and it just flew.
As a rocket engineer, who designs rockets, Musk is probably replaceable. As a rocket company engineer, who designs rocket companies, he is not.
For over 50 years, the US space industry stagnated precisely for the lack of a Musk-like figure, who could not only create an effective aerospace company, but who could thereby demonstrate what an effective aerospace company looks like.
Which makes it much more difficult for fraudsters to get away with fraud.
I'm envious. I wish I could have worked for an aerospace company instead of a fraud company.
But what's more important is that some of us now no longer have to play "who gets the most cheese?".
They get to play "What can we build today?".
Have you ever wondered why catching Starship is so much harder than catching Super Heavy, even though they both return to the same launch tower? 🔥
💡 The answer comes down to four major engineering challenges that make Starship’s recovery fundamentally different 👇
⚡ The first problem is shedding an enormous amount of energy
At hot staging, Super Heavy was traveling at roughly 5,800 km/h (3,600 mph). Starship, meanwhile, had to return from orbital speeds of roughly 27,500 km/h (17,100 mph).
That’s nearly five times faster. Because kinetic energy increases with the square of velocity, each kilogram of Starship at orbital speed carries roughly 22 times the kinetic energy of a kilogram of Super Heavy at separation.
That works out to nearly 30 million joules per kilogram.
Almost all of that energy has to be dissipated before Starship can safely reach the tower.
One option would be to use rocket propulsion to cancel out its orbital velocity. However, scrubbing off roughly 7.7 km/s of speed using Raptor engines alone would require a massive amount of propellant.
The rocket equation makes this approach especially punishing. Every extra kilogram of landing propellant has to be hauled into orbit first, dramatically increasing the fuel needed at liftoff.
Instead, SpaceX uses Earth’s atmosphere to bleed off almost all of Starship’s orbital velocity.
That saves an extraordinary amount of propellant, though it introduces another major hurdle.
🔥 Surviving extreme atmospheric heating comes next
As Starship plunges into the atmosphere at hypersonic speeds, powerful shock waves compress and heat the surrounding air to blistering temperatures.
Thousands of ceramic heat-shield tiles protect its stainless-steel structure from the intense heat.
Simply keeping those tiles attached isn’t enough, though. Superheated gas can seep into the gaps between them, risking damage to the underlying structure.
For Flight 15, SpaceX carried forward upgrades from Flight 14, including extra tile-retention hardware in vulnerable areas and redesigned curved tiles that reduce heating between adjacent tiles.
Because Super Heavy never reaches orbital velocity, its return doesn’t expose it to anything close to the same thermal stress.
And there’s another complication.
A heat shield that survives one reentry isn’t necessarily ready to fly again. Starship’s thermal protection system must eventually withstand repeated missions without major repairs or refurbishment.
That’s critical, since rapid reusability is one of the core reasons SpaceX designed Starship to be caught rather than discarded.
Yet surviving reentry still isn’t enough. Starship also has to arrive at the exact right spot.
🎯 Hitting a precise target from orbit raises the stakes even further
About 56 minutes before the catch, a single Raptor engine fired for just 10 seconds to initiate Starship’s return.
That brief deorbit burn wasn’t meant to eliminate orbital velocity. Instead, it nudged the spacecraft’s trajectory so atmospheric drag could take over the much larger job of slowing it down.
Starship then entered the upper atmosphere thousands of kilometers away over the Pacific Ocean.
From there, its four aerodynamic flaps controlled its orientation and fine-tuned the forces acting on the vehicle.
Super Heavy uses grid fins to steer during its return, but Starship’s flaps have to handle a far wider range of flight conditions.
As Starship decelerates from hypersonic through supersonic and down to subsonic speeds, its guidance system continuously adjusts its attitude and flight path.
The challenge is maintaining the right balance of altitude, speed, and trajectory while aerodynamic forces shift dramatically throughout reentry.
For Flight 15, the return corridor crossed sparsely populated parts of northern Mexico and South Texas.
Starship maneuvered around major population centers in the Rio Grande Valley, passed north of South Padre Island, headed offshore over the Gulf of Mexico, and made its final turn toward Pad 2 at Starbase.
Think about the guidance precision required.
A spacecraft returning from orbit must dissipate nearly all of its immense kinetic energy while threading a carefully controlled trajectory straight back to a single launch tower.
It also has to arrive at a low enough speed, in the proper orientation, and with enough propellant remaining to pull off the final maneuver.
That brings us to the most visually spectacular part of the entire recovery.
🚀 Executing the landing flip with pinpoint precision is the final test
Super Heavy approaches the launch tower already vertical, using its grid fins and Raptor engines to guide its final descent.
Starship comes in belly-first.
It has to rapidly swing its massive frame upright, stabilize its attitude, and align itself precisely between the tower’s mechanical arms.
SpaceX’s official Flight 15 timeline allotted just 18 seconds between the start of Starship’s landing burn and the catch.
During those 18 seconds, three sea-level Raptor engines ignite to kick off the maneuver.
The spacecraft rotates upright while its flight computers constantly regulate engine thrust and vehicle attitude.
The engines then step down from three to two, giving the vehicle finer control as it nears the chopsticks.
Finally, it transitions to a single engine for the last moments of descent before shutting down as the tower takes its weight.
Consider the coordination involved.
The guidance system must balance translation, rotation, thrust, and descent rate while positioning a massive spacecraft within a narrow capture envelope.
Unlike a conventional vertical rocket landing, Starship must first transition from an aerodynamically controlled, belly-first glide to an engine-powered vertical approach.
All of this unfolds within seconds, right after the vehicle has completed one of the most punishing atmospheric reentries ever attempted by a reusable spacecraft.
🧠 What makes SpaceX’s approach so ingenious is how seamlessly the entire recovery system works together
Trying to solve every phase of the return with rocket propulsion would carry a crushing propellant penalty.
Instead, SpaceX delegates the toughest parts of the recovery to systems purpose-built to handle them.
The atmosphere dissipates most of the orbital kinetic energy without consuming onboard propellant.
The heat shield protects Starship from the intense thermal loads of that deceleration.
The four aerodynamic flaps provide the control authority needed to manage its attitude, energy, and trajectory throughout atmospheric flight.
And the Raptor engines take over once aerodynamic control alone can no longer deliver the precision needed for the final catch.
That division of labor allows SpaceX to recover an orbital spacecraft without carrying enough fuel for an all-propulsive return.
Yet the engineering trade-off goes deeper than saving propellant.
SpaceX is shifting much of the cost of returning from orbit away from expendable propellant and toward reusable hardware built to endure extreme thermal and mechanical stress.
The potential economic payoff is enormous.
Cutting back on the propellant reserved for recovery frees up more of Starship’s mass capacity for useful payload. However, that advantage only becomes truly meaningful if the heat shield, flaps, engines, and other systems can survive repeated flights without extensive refurbishment.
That’s why catching Ship 42 is such a landmark milestone.
SpaceX has proven that its thermal protection, aerodynamic controls, propulsion, and guidance systems can work in concert to recover an orbital upper stage right at the launch site.
The next engineering hurdle is proving that the same spacecraft can do it repeatedly.
If SpaceX can bring Starship home, inspect it, refuel it, and launch it again without rebuilding major systems, the economics of getting payload to orbit could change dramatically.
That first catch proves the recovery architecture works. Repeated reflights will determine just how much of its economic potential SpaceX can unlock.
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 https://t.co/YOtzND0Jqk.
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.
The spy novel is the sideshow. I read four years of Nvidia's SEC filings and wrote it all down: the numbers, the footnotes, the law, the CFO, the auditor, the SEC, DOJ and PCAOB. The fullest account I know of the question nobody is asking. https://t.co/ysxBFvXjqu
Or could be circumscribed restatement of admission in his letter to employees:
"I was unfaithful to my wife. My actions hurt her and our children. There is no excuse for what I did. I am deeply sorry."
https://t.co/OSLu4pMqf2
Despite the often described as insatiable demand in the US, $NVDA could only maintain its revenue growth by selling to SE Asia for use in China (via physical smuggling and remote access) because US purchases are gated by availability of electricity to power immediate deployments.
“I am deeply sorry for what I’ve done.”
Reported text from Jensen Huang to @altcap Brad Gerstner. Via CNBC.
Sounds more expansive than apology for inappropriate romantic relationship. Hope message doesn’t betray guilt over broader misconduct.
“I'm sorry for my mistake”: acknowledgement of discrete errors
or
“I'm sorry for having an affair”: express acknowledgment of specific misconduct
vs
“I'm sorry for what I've done": unqualified without definition is psychologically evocative because it leaves the scope of the wrongdoing unspecified
The spy novel is the sideshow. I read four years of Nvidia's SEC filings and wrote it all down: the numbers, the footnotes, the law, the CFO, the auditor, the SEC, DOJ and PCAOB. The fullest account I know of the question nobody is asking. https://t.co/ysxBFvXjqu
Despite the often described as insatiable demand in the US, $NVDA could only maintain its revenue growth by selling to SE Asia for use in China (via physical smuggling and remote access) because US purchases are gated by availability of electricity to power immediate deployments.
𝗦𝗲𝗺𝗶𝗔𝗻𝗮𝗹𝘆𝘀𝗶𝘀 𝗷𝘂𝘀𝘁 𝗰𝘂𝘁 𝗶𝘁𝘀 𝗔𝗜 𝗰𝗮𝗽𝗲𝘅 𝗳𝗼𝗿𝗲𝗰𝗮𝘀𝘁, 𝗮𝗻𝗱 𝗡𝘃𝗶𝗱𝗶𝗮 𝘁𝗮𝗸𝗲𝘀 𝘁𝗵𝗲 𝗯𝗿𝘂𝗻𝘁 $NVDA
In its Nov 13th institutional model update, SemiAnalysis lowered its forecasts for AI accelerator shipments, spending, HBM and power demand, with 2028 to 2030 cut the hardest.
𝗧𝗵𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 (𝗴𝗹𝗼𝗯𝗮𝗹 𝗔𝗜 𝗮𝗰𝗰𝗲𝗹𝗲𝗿𝗮𝘁𝗼𝗿 𝗰𝗮𝗽𝗲𝘅)
• 2026: ~$392B, roughly unchanged
• 2027: $770B to $627B (-19%)
• 2028: $1.054T to $758B (-28%)
• 2029: $1.352T to $796B (-41%)
• 2030: $1.531T to $923B (-40%)
Also cut: 2028 HBM demand by 30% and 2030 accelerator power demand by 34%.
𝗪𝗵𝘆 𝘁𝗵𝗲𝘆 𝗰𝘂𝘁
• Frontier lab net new ARR growth slowed in Q3, so they reset expected token revenue to what can sustainably pay for compute
• Capital markets have tightened. Debt and equity investors are pulling back from AI capex that lacks hyperscaler commitments
• The midterm results delayed 5.8 GW of US data center supply planned for 2H27 to 1H28, though they stress the first two reasons matter more
𝗪𝗵𝘆 𝗡𝘃𝗶𝗱𝗶𝗮 𝗯𝗲𝗮𝗿𝘀 𝘁𝗵𝗲 𝗯𝗿𝘂𝗻𝘁
The cuts don't land evenly. Customers keep buying the best performance per TCO and per watt, and the weakest platforms on those metrics take disproportionate cuts. By their framework, that is Nvidia:
• Nvidia accounts for 66% of the cut in cumulative 2026 to 2030 accelerator wafer area
• Its 2030 shipment forecast on a compute revenue basis is cut 66%
• Lower ASP assumptions for Feynman and beyond as competition rises
• At Anthropic, Nvidia GPUs sit at the margin next to TPU, Trainium and AMD. OpenAI is leaning on AMD and its own ASICs
• Highest exposure to neoclouds, especially the smaller long tail with more financing risk
𝗠𝘆 𝘁𝗮𝗸𝗲
Read this carefully. It is not a call that AI demand is dying. Even on the new numbers, accelerator capex more than doubles from $392B in 2026 to $923B in 2030.
What it says is that the buildout is now limited by financing and revenue, not chips. That ties directly to what I have been saying about liquidity and the Anthropic raise pulling capital out of the market.
The winners from here are whoever has the best cost per token and the strongest balance sheet behind the order. Coming the same week Jensen left and the Treasury Secretary questioned Nvidia's revenue quality, this is one more weight on Tuesday's guide.
Translated excerpt attached. Credit @dmjk001 for original excerpt.