Grant (@glbeaty) header
Grant (@glbeaty) profile picture
Joined May 2015
Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)Photo by Grant (@glbeaty)
Downward pressure on GPU rates, as Chinese-rented GPUs get put back on the market. Upward pressure on physical GPUs, which can be smuggled into China and served there.
Datasection (3905.T) just cut FY revenue guidance by 89%, from ¥162B to ¥17B, and the stock went limit-down. This is the "Japanese neocloud" that the FT reported is effectively a Tencent datacenter: Chinese offtake, Nvidia Blackwell GPUs, Japanese flag. (1/8)🧵
1
2
859
Seems like Bessent is walking back the spy narrative. If I had to come up with one headline for this $NVDA scandal it'd be: Failed Macro Trader, 64, Still Doesn't Realize Sanctions Don't Stop Trade The market also falls for this, every damn time. youtube.com/watch?v=YmraKlyE…
7
838
GPU rates for the past 4 weeks. If AI crashing reduced demand through VC funding fear, it's not showing up yet. There could also be a counteracting force, with lower capex estimates prompting more long-term leases. We're getting a test of one of the last bear theses remaining.
1
1
8
1,299
Thoughts on the results: I don't know how to interpret "they're not", as this implies no bubble or bear case, at least for the hyperscalers. VC-subsidization seems flimsy: could be wasteful, or a startup could unveil the ChatGPT moment for robotics tomorrow. No way to know?
AI bears, why are GPU prices wrong? See reply for the forward curves. Everyone knows about token price wars and the amount of capacity coming online, so I don't think supply is going to blindside anyone (unless #2). If none of the below, please reply with your hypothesis.
8%Compute speculation
18%Dumb compute buyers
26%VC-subsidized training
48%They're not
85 votes • Final results
2
6
1,223
Dumb buyers: consider the costs of over v. under-building compute. A big shortage pushing GPU rates to the wages they replace could see rates 100x! Erring on too much is rational for companies & the economy. Still, dumb buyers may not be tallying model efficiency improvements.
1
364
Lastly, I haven't seen evidence of compute speculation independent of hedging or risk aversion. Doesn't mean it's not happening. It's hard to over-emphasize how costly under-building could be, or how much efficiency gains overwhelm physical capex.
286
In other news, $NVDA shareholders are not considering keeping their board of directors 😂 If Jensen didn't break serious laws, I hope they have good personal insurance! Their $900M of aggregate D&O ain't gunna cut it. This all smells of hysteria. cnbc.com/2026/11/16/nvidia-c…
1
8
883
The $4T question is, did $NVDA break sanctions laws by selling chips to China? Why, when they can sell chips to someone who serves them to China? It's called the remote access loophole for a reason. Serving China chips made by TSMC is aligned with Taiwan, but not Washington.
1
9
1,427
I'd bet no serious laws were broken here, so there's a good chance Jensen returns. Maybe the pressure to fire him was typical Trumpian bargaining: open with the most extreme position possible? This might've all been about increasing Jensen's alignment with Washington.
1
3
662
Jensen wants (from most to least): Adultery 🤷‍♂️ To keep his job Not to go to jail $NVDA revenues AI model competition Trump wants: stonks go up U.S. model dominance Taiwan wants: China addicted to their chips (maybe) Western defense (definitely) Using the loophole made sense.
1
1
438
AI bears, why are GPU prices wrong? See reply for the forward curves. Everyone knows about token price wars and the amount of capacity coming online, so I don't think supply is going to blindside anyone (unless #2). If none of the below, please reply with your hypothesis.
8%Compute speculation
18%Dumb compute buyers
26%VC-subsidized training
48%They're not
85 votes • Final results
1
1
1
2
2,991
Neocloud forward curve settlement prices going back to July, from Silicon Data. Hyperscaler prices are considerably higher, especially for shorter durations. I don't know what the dip at 2 years is about, but 1 or 3 year terms seem to be the most common. No contango in sight.
1
1
4
2,617
This works out to 3 years of a B200 costing $169k, creating about $144k of cash (minus power, labor) for the owner. The fixed cost (GPU, networking, server, datacenter, land, appeasing NIMBYs, etc.) should be ~$100k / B200. Not nearly as good as hyperscaler economics.
597
Where were the signs? 😂 "Exponential growth" is often growth of many (sub-)linear growth drivers compounding. Ironically the truth seems more bullish semis? Transistor size was the exception: a halving meant ¼ die area, ½ voltage, ½ switching time, & ⅛ energy / switch.
The following is from Leopold Ashenbrenner’s Situational Awareness on AI. It is inappropriate to make the claim “it just requires believing in straight lines on a graph” with logarithmic scale. The more appropriate statement is “it requires believing in exponential growth”. This is how most forecasting errors happen. Complex systems have non-linearities and boundary conditions. What is the point that you run out of power? That there is insufficient new data for novel learnings (creating new synthetic data does not address distributional gaps). The next decade will be amazing. But straight lines are deceptive to the complexity at hand.
3
1,107
I don't have a high opinion of this admin, but clearly Jensen (who is Taiwanese) lobbying to sell chips to China was always sus? My headcanon is Jensen promised an undetectable backdoor, then was thrown under the bus when the Chinese discovered it. nvidianews.nvidia.com/news/a…
1
1
9
1,393
So this wasn't my imaingination? I don't think it's about gross margins: $ANT's are great & the market mostly looks at ARR. This calls into question all closed model benchmarks without samples over time. Are they improving as quickly as thought, or are older models held back?
After Anthropic made Fable 5 permanently available in subscription plans, I noticed a large drop in performance. The model felt dumber, and I couldn't explain why. Measured five different ways, August delivered dramatically fewer thinking tokens than July.
1
4
1,752
Is the market seriously thinking U.S. politics will seriously disrupt the global demand for AI compute? Or is this a scenario where sellers believe others will sell, so they're selling first?
18%Yes, Blue Fear is real
30%No, wants to sell first
42%No, but pins pop bubbles
10%Something else?
100 votes • Final results
1
4
1,167
As unlikely as it is, U.S. stopping future datacenter building probably has the same long-term effect on the supply of AI as sanctions on global commodities that can be loaded on boats: None at all. There are 50 states and 195 countries to chose from.
3
12
1,016
$ANT Effective altruists gunna altruize 🤷‍♂️ Say what you will about Dario, but he seems like a good guy.
1
9
4,140