Mach

Aerospace engineer | Aircraft engines by day, hooked on markets by night | Long AI since GPT-2

[Mach @MachStack]

AMD beat NVIDIA by 5.9% in published peak HBM4 bandwidth. It did so with 50% more stacks.

AMD won the headline. NVIDIA may have built the better configuration for an HBM shortage.

Why AMD's HBM4 advantage could backfire: https://t.co/KYRCE5PmAZ

[Mach @MachStack]

@KobeissiLetter that's basically why there are so many index fund guys, you get great long-term returns and don't have to spend your life worrying about what to buy or sell

[Mach @MachStack]

@StockSavvyShay @Fiscal_ai the clear downside here is margins though, Nvidia $NVDA expects them to come down over the next few quarters as memory gets more expensive

so yeah, the commitments look bullish for demand, but locking in that much supply isn't free

[Mach @MachStack]

@DanDr1s I don't know if METR will publish a time horizon for Astra when it drops, but I'm convinced we're already talking about multi-day work horizons

[Mach @MachStack]

if you look at Anthropic they seem to be going the other way, Haiku has basically been left behind while they keep pushing Sonnet and Opus

I think that makes more sense as a business

US labs are going to have a really hard time competing with Chinese models on price, I'd rather focus on the mid and high end where capability matters more and there's more revenue to capture

[Mach @MachStack]

It's been almost 3 months since we had a clear move at the AI frontier

at the speed this field moves, that feels like forever

looks like September is finally going to break the streak 🤞

[Mach @MachStack]

@kimmonismus it hasn't even been 2 years since reasoning models started showing up

we're still in the baby stage

[Mach @MachStack]

There are a few ways to deal with this

One is cutting spending aggressively and risking a recession and higher unemployment. Another is massively increasing productivity and growing your way out of it

CBO estimates that every extra 0.1pp of productivity growth can cut deficits by roughly $300B over a decade

AI + robots might have arrived at exactly the right time

[Mach @MachStack]

@mark_k Not a model, but I'd pay close attention to OpenAI DevDay

with all the rumors around persistent agents, this could be their most impactful DevDay yet

[Mach @MachStack]

@Nairebis @PeterDiamandis ah and on Starlink, I don't think that goes against my point

it had a real edge, especially connectivity in places where terrestrial infrastructure made no economic sense

that's basically my point

[Mach @MachStack]

Read what I wrote again, I never asked you to predict the future, I asked what exactly is supposed to create this new economy. Explaining the mechanism and predicting the outcome are two completely different things

With orbital data centers, the problems we have on Earth honestly look much easier to solve than the ones you create by moving compute into space

The power isn't really "free" when you have to launch the chips, solar arrays, radiators and everything else needed to keep them running. Cooling doesn't disappear either, you just replace terrestrial cooling with huge radiators that also have to be launched

Then there's maintenance and depreciation, SpaceX itself says it doesn't expect to repair processors in orbit, so when something fails or becomes obsolete you're replacing capacity entirely. That's a pretty big issue for AI hardware that already depreciates insanely fast

And I wouldn't assume the permitting advantage stays forever either, if this actually scales to hundreds of thousands of satellites, regulation can tighten as this scales just as launch costs can keep falling

Space tourism could become a big market, sure, but to me that's closer to deep sea tourism than a new productivity revolution. I really don't see how that ends up massively moving global GDP

Even SpaceX itself treats tourism, in-orbit manufacturing, Moon/Mars manufacturing and asteroid mining as unquantified future markets

I think the space economy can become absolutely massive, far bigger than it is today

I just don't see anything here that makes me think it's going to make global GDP explode

[Mach @MachStack]

@KobeissiLetter one reason this hasn't turned into a real correction is software finally coming back to life

semis have had some brutal days lately, but software has been offsetting a lot of that weakness

[Mach @MachStack]

I'm not asking anyone to predict exactly what happens

I'm asking what the actual economic edge is. If we're talking about trillions in new GDP, there has to be some reason doing these things in space eventually becomes better than doing them on Earth

that's the part I'm trying to understand

[Mach @MachStack]

cell phones had a clear edge over landlines because you could take them anywhere

that's exactly what I'm asking about space in general, what unique advantage does it unlock that's valuable enough to overcome the extra cost?

compute, mining, energy... what's the equivalent of mobility here?

[Mach @MachStack]

@StockMKTNewz @Fiscal_ai and if you look at profits the gap is even crazier

Nvidia $NVDA made 8.3x AMD's revenue last quarter, but 26x its net income

[Mach @MachStack]

I get the long-term argument, but I'm not sure America is a great comparison

America gave access to huge amounts of productive land and resources that Europe either didn't have or couldn't produce at anything close to the same scale. There was a clear economic reason to exploit that new territory

a lot of the space economy being pitched today feels different, you're often taking things we already do on Earth, compute, mining, energy, and trying to make them work somewhere with a massive cost disadvantage

maybe those costs collapse over time, but that's the part I'm trying to understand, what does space eventually do so much better than Earth that it creates an entirely new economy?

[Mach @MachStack]

Even the best AI agent today gets a 30% resolution rate on expert-built scientific workflows

Terminal-Bench-Science tested 70 tasks across life, physical, Earth, mathematical and engineering sciences, with 3 independent runs per task

Claude Opus 5 + Claude Code: 30.0%
GPT-5.6 Sol + Codex: 22.4%
Claude Fable 5 + Claude Code: 21.4%
GLM 5.3 + Claude Code: 8.1%

the cost gap is pretty crazy too

Sol basically matched Fable for $4.2K vs $14.2K, while actually using more tokens, 8.4B vs 6.4B
these aren't exam questions either, the agent has to actually do the work: analyses, simulations, proofs, code, etc. and each task has its own verifier

this benchmark, just like Humanity's Last Exam or FrontierMath, was specifically built to be hard for frontier models

so current models can have a much bigger impact today than a 30% score here might make you think

[Mach @MachStack]

@Dr_Singularity Welinder clarified this later, he thinks AGI will come pretty fast

the few years are for the physical constraints around real-time inference and cheap edge hardware, not AGI itself

[Mach @MachStack]

@sattyyouneed For the same reason people still buy Nvidia GPUs when cheaper alternatives exist

Being cheaper and good enough isn't the same as being the best. The US still holds the lead at the frontier

[Mach @MachStack]

@TimJayas Yeah, there have been some rumors about the new Fable but so far everything that's come out makes it look like it'll be below Astra

we'll see

[Mach @MachStack]

Most of that increase is related to memory procurement, Nvidia $NVDA is basically locking in supply years ahead

even with memory prices going crazy they only expect gross margins to go from 75% to ~74% next quarter, before bottoming around 71-72%

looks a lot more like preparing for massive demand than a bearish signal to me