The reason every forecast including Anthropic's own keeps being wrong in the same direction is that the existing frameworks are built for single-loop technologies. Electricity had demand-side network effects, more people on the grid meant more appliances got built which drove more demand. AI tokens have something with no precedent: simultaneous demand-side AND supply-side network effects. More usage creates more use cases (demand loop). But more usage also generates training signal and fine-tuning data that makes the model better (supply loop). The consumption of the product is itself the R&D that improves the product.
No previous technology had both loops running at once. Electricity doesnt improve when you consume more of it. But every token consumed makes the next token more valuable because the model underneath has improved from the interaction. Thats why the 80x number keeps surprising even the people building it, theyre forecasting with single-loop frameworks against a double-loop system and the error compounds in every cycle. The electricity analogy is directionally right but it actually understates the dynamic, because electricity never had the property where the act of consumption was simultaneously an act of production.
Well stated,. It is indeed the innermost loop. The tool begets more use cases. More use cases get the tool to be better. The tool's next phase of improvement begets ANOTHER level of use cases. And on and on we go, and where it stops NOBODY knows.
Really nice read. super interesting that grok was only using 11% of Colossus. The deal ended up working out well for them, but i still didn't realize how low the usage was.
I've been going through a few recent podcast transcripts from people who are either inside or very close to Anthropic - the CFO, the founding investor, and a couple of VCs tracking token consumption closely. A few things stuck out.
The capex math is worth sitting with.
Compute commitments at Anthropic reportedly run at approximately 3-4x current revenue. When revenue is growing 10x per year, the company is committing ~$30B in capex for every $1B in current revenue. That's why the $50B raise at a $900B valuation isn't irrational - in that structure, there's no such thing as too much capital on the balance sheet.
The Colossus deal reduces the near-term constraint. It doesn't fix what Opus 4.7 revealed about competitive dynamics.
Opus 4.7 had compute gating applied to it - reduced thinking time to manage GPU scarcity before the deal closed. Enough developers noticed the quality degradation that many rolled back to Claude 4.6. That happened right as OpenAI's GPT 5.5 (new base model, "Spud") was getting strong developer reviews, and Codex was gaining coding token share.
These dynamic of competition and compute constraints can lead to interesting times in the next 6 months.
One “book feedback@ - “create” beef easily and the next part of book about managing stroke and health - maybe AI should be the true guide for humans ??
Grok will SpaceXA1. Collusus and Collusus II have compute power. You didn't mention Cursor. SpaceX is positioned well and will make a great investment.
Peter, one of things not mentioned in the Metatrends newsletter is the buzz around the combination of Codex and ChatGPT 5.5. Serious users are raving about not only its coding ability but the complete integration of Agents and Skills in the UI. Claude looks disaggregated in comparison. This new framework is OpenAI's attempt, in my view, to shoehorn its way into the corporate space. But even more significant, it gives a one person entrepreneurial shop the ability to cook up customized applications quickly and efficiently. See the video from AWG's pal Alex Finn, who was on the Moonshots podcast.
The reason every forecast including Anthropic's own keeps being wrong in the same direction is that the existing frameworks are built for single-loop technologies. Electricity had demand-side network effects, more people on the grid meant more appliances got built which drove more demand. AI tokens have something with no precedent: simultaneous demand-side AND supply-side network effects. More usage creates more use cases (demand loop). But more usage also generates training signal and fine-tuning data that makes the model better (supply loop). The consumption of the product is itself the R&D that improves the product.
No previous technology had both loops running at once. Electricity doesnt improve when you consume more of it. But every token consumed makes the next token more valuable because the model underneath has improved from the interaction. Thats why the 80x number keeps surprising even the people building it, theyre forecasting with single-loop frameworks against a double-loop system and the error compounds in every cycle. The electricity analogy is directionally right but it actually understates the dynamic, because electricity never had the property where the act of consumption was simultaneously an act of production.
Well stated,. It is indeed the innermost loop. The tool begets more use cases. More use cases get the tool to be better. The tool's next phase of improvement begets ANOTHER level of use cases. And on and on we go, and where it stops NOBODY knows.
Really nice read. super interesting that grok was only using 11% of Colossus. The deal ended up working out well for them, but i still didn't realize how low the usage was.
Thus article and knowledge is blowing my mind. I love it and I thank you for your optimism and vision of helping us see the future better.
I've been going through a few recent podcast transcripts from people who are either inside or very close to Anthropic - the CFO, the founding investor, and a couple of VCs tracking token consumption closely. A few things stuck out.
The capex math is worth sitting with.
Compute commitments at Anthropic reportedly run at approximately 3-4x current revenue. When revenue is growing 10x per year, the company is committing ~$30B in capex for every $1B in current revenue. That's why the $50B raise at a $900B valuation isn't irrational - in that structure, there's no such thing as too much capital on the balance sheet.
The Colossus deal reduces the near-term constraint. It doesn't fix what Opus 4.7 revealed about competitive dynamics.
Opus 4.7 had compute gating applied to it - reduced thinking time to manage GPU scarcity before the deal closed. Enough developers noticed the quality degradation that many rolled back to Claude 4.6. That happened right as OpenAI's GPT 5.5 (new base model, "Spud") was getting strong developer reviews, and Codex was gaining coding token share.
These dynamic of competition and compute constraints can lead to interesting times in the next 6 months.
We just need more creative solutions to meet the moment.
It is no longer necessary to possess subject matter expertise to contribute to any field. Transduction.
https://substack.com/@johnbaptist42/note/c-260356275
One “book feedback@ - “create” beef easily and the next part of book about managing stroke and health - maybe AI should be the true guide for humans ??
The post uses stale numbers for Anthropic. Their ARR is over $44 billion now. Not $30B.
$45 billion in projected annualized revenue
Up from $9 billion at the end of 2025
5x growth in roughly 6 months
45x revenue increase in 17 months.
https://www.the-ai-corner.com/p/anthropic-1-trillion-valuation-dario-amodei-2026-breakdown
Grok will SpaceXA1. Collusus and Collusus II have compute power. You didn't mention Cursor. SpaceX is positioned well and will make a great investment.
Great post. Thanks.
Peter, one of things not mentioned in the Metatrends newsletter is the buzz around the combination of Codex and ChatGPT 5.5. Serious users are raving about not only its coding ability but the complete integration of Agents and Skills in the UI. Claude looks disaggregated in comparison. This new framework is OpenAI's attempt, in my view, to shoehorn its way into the corporate space. But even more significant, it gives a one person entrepreneurial shop the ability to cook up customized applications quickly and efficiently. See the video from AWG's pal Alex Finn, who was on the Moonshots podcast.