chris

JUST IN: 🇸🇬 Singapore's Bitdeer becomes the largest Bitcoin miner by total capacity Total hash rate under management: 71 EH/s 🔥

We are proud to have Bitdeer AI as a NVIDIA Nemotron 3 Ultra launch partner, bringing ready-to-run access with NemoClaw on its secure cloud platform. Build your AI Agents with Bitdeer AI from Day 0.

Hyperscalers will be the ultimate winners of this. Here is how this will likely play out: As model capabilities plateau, more of the spending will shift to cheaper models. This will hit frontier labs like OpenAI and Anthropic. Enterprises will use Chinese models through American cloud providers. These cloud providers will make the margins. Frontier labs will stop attracting investment as their economics won’t work, thus frontier progress will stall. Hyperscalers like $MSFT, $META, $GOOG, $AMZN will make equivalent of those cheap AI models by distilling SOTAs in the same way. Enterprises will prefer American models if the cost/performance is the same. Thus, hyperscalers will become vertically integrated AI giants, expanding their margins on every inference token. Infrastructure wins.

JUST IN: 🇺🇸🇨🇳 American companies now use Chinese AI models more than US-made ones.

$SKHY 4.87x. You gotta be kidding me. Even lower than $MU at 5.56x. At these levels, they can buy back the entire company in just ~4 years. Shortage until 2030. Pure madness.

China's new frontier AI lab 'Moonshot' released its flagship 'Kimi K3' model today, which is initially benchmarking as competitive with Anthropic, OpenAI & xAI's newest models. 'Moonshot' is valued at just $31.5 billion. Anthropic & OpenAI are prospectively being valued north of $1 trillion, each... For those who are wondering why this rattled AI-related markets today, even in spite of a massive unwind over the past month, the logic is straightforward: - Hundreds of billions of dollars of capex & spending commitments are intimately tied to the U.S. frontier labs (Anthropic, OpenAI & xAI namely) - Capex has been the lifeblood of the AI trade. The U.S. semiconductor complex and the 2nd order datacenter buildout beneficiaries, including neoclouds & power companies & construction companies and many others at the periphery, have seen their stocks rerate higher on the anticipation that this capex will flow to their top & bottom lines. - Cheap, but competitive, Chinese models trained at a fraction of the cost of their U.S. peer-models, produced by firms who are valued at a fraction of their U.S. peers, who are offering usage/tokens at a fraction of the cost, pose the risk of "commoditization of intelligence". While this is certainly a tailwind for the proliferation of AI usage across industries, it poses a risk to the U.S. model of spending & committing massive amounts of capital to train & deploy new AI models. - If open-weight, cheaper, competitive models lead to significant compression in the pricing power for U.S. frontier labs, what impact does this have on the enormous spending commitments they have made in the near-future, and what 2nd order impacts does that have on the companies that are due to receive, benefit from, or deploy that capex? - As a hedge to this argument, it is also very true that increased usage of AI will require even more spending on inference capacity (Kimi K3 is a very big model at 2-3 trillion parameters, the biggest to ever come out of China), which could offset some of the compute-related risk in the training portion of this equation. The question is how does the calculus change if the training front-end of this equation has to be reworked entirely. It may also plant seeds for the conversation of improving inference efficiency through innovation (note that the efficiency through innovation conversation had already begun to emerge on an entirely different front, with memory stocks, several weeks ago). - When we are in momentum-driven periods with inordinate amounts of spending -- spending that is beginning to jeopardize the profitability of the most valuable companies in history, sometimes all it takes is a small seed of doubt to change the expectations & dynamics & internal balance-sheet considerations, and that seed of doubt can have domino effects. - While eerily similar to 'Deepseek', that model was not as competitive, and also that conversation happened at a time when Nvidia was the lifeblood of the AI trade, and the real question at the time was moreso whether or not they were using Nvidia's most cutting-edge chips, and had little to do with the commoditization of intelligence issue. Shortly after the "Deepseek moment", tariffs emerged and overshadowed it as a much more real & present threat to the markets & the economy. We moved on rather quickly, and to an extent, began whistling past the graveyard, proverbially speaking. To be abundantly clear before the AI bulls unleash on me in the comments: I'm not an AI bear, at all. I've been long AI stocks for almost the entirety of the past 3 years. I am still long many of them. This is just legitimate food for thought about the capital-equation for the U.S. AI industry. Digestion is warranted. I'm still navigating this issue, and I'm more than happy to hear & debate counterpoints on the implications and possibilities from any & all.

Gavin Baker: “It would be bullish for AI infrastructure if the model market shifts away from expensive flagship models to cheaper models. This would shift margins from AI labs to cloud providers.” This is bound to happen. Look at the chart below. It looks like Opus 4.7-4.8 class is the sweet spot for most of the current workloads. Apparently, paying more for Fable 5 doesn’t bring big enough performance jump to justify higher prices per token, as it captures only a small portion of the spending. This means that, when cheaper models categorically achieve Opus 4.8 level performance, we can expect spending to shift toward them. When this happens, token prices will collapse, enterprises will use more tokens, run more experiments, and become more productive and innovative. There will be many of these models. Meaning, they’ll essentially be commoditised so those who make the margins will be the cloud providers that run these models at scale. Long AI infrastructure. $MSFT $AMZN $GOOG $NBIS $CRWV