OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze

TAKEAWAYS

My biggest takeaways from this week's All-In

1. 01

   OpenAI's miss is a power problem, not a demand problem.

   Chamath reads the WSJ report on OpenAI missing its billion-weekly-user and revenue targets as a supply story: nobody in this market is short of demand, they are short of the electricity needed to generate tokens. Backlogs in power and components hurt OpenAI and Anthropic most and hand leverage to the hyperscalers, who end up trading compute for equity and control. His verdict on the misses: 'It is entirely 100 percent due to the supply of the power necessary to generate the output token.'

   “Everything in this market is power constrained. The reason that these folks may miss a number or a forecast have nothing to do with demand. It is entirely 100 percent due to the supply of the power necessary to generate the output token.”

2. 02

   There are no dark GPUs, so this is not Cisco in 2000.

   Sacks rejects the dark-fiber analogy Jason raises against roughly $725 billion of 2026 hyperscaler CapEx. In 2000 the fiber went dark and Cisco needed 25 years to regain its peak; today every GPU that comes online is consumed by demand for tokens, and that demand is what pulls the investment forward. He counts the simultaneous Google, Microsoft, Amazon and Meta beats as the moment the bull thesis got validated in a single afternoon.

   “There's no dark GPUs today, as Brad Gerstner likes to say.”

3. 03

   AI is now synonymous with the American economy, and it won't bring the job loss the doomers predict.

   Sacks puts AI at about 75% of GDP growth last quarter, with CapEx alone above 2% of GDP. That makes halting AI equivalent to halting the economy, which is why he thinks the political case against it collapses. On the labor side he sees a productivity boom rather than mass displacement, and treats 'eliminating all software developers' as the peak of inflated expectations that has already passed.

   “I don't think there's going to be this huge job loss associated with this productivity boom that we're going to get.”

4. 04

   In five years the hyperscalers look like bulky industrials. Follow the dollars to their suppliers.

   Chamath sees a structural shift from asset-light to asset-heavy capital markets. Around $700 billion a year of CapEx, collapsing free cash flow (Amazon down 97%) and heavy leverage turn Google, Microsoft, Amazon and Meta into bulky industrial businesses with no obvious valuation case at the end of it. His example of what they are willing to pay: Microsoft's Three Mile Island power agreement ran at more than twice the spot rate. The trade is to buy the underpriced asset-heavy suppliers the trillion dollars flows through.

   “they're going to look like this big bulky industrial business in five years. I'm not sure that there's a good valuation case to be made at that point.”

5. 05

   Any mature market ends up 4-2-1, and AI will be two of them: consumer and enterprise.

   Friedberg applies BCG's Rule of Three: a stable competitive market settles into a 4-to-2-to-1 share ratio among three players covering about 90% of the market. He splits AI into two such markets. Consumer trends toward a ChatGPT-versus-Google fight for first and second with Anthropic third; enterprise is led by Anthropic or Google. OpenAI's 900 million weekly users still dwarf Claude's, which he puts at sub-100 million.

   “any stable, mature, competitive market evolves to a 4 to 2 to 1 ratio of market share for basically 90% of the market”

6. 06

   Sam will be right about his compute bet for the wrong reason: coding, not consumer.

   Sacks argues the $600 billion in compute commitments looked reckless against a consumer miss, but coding became the all-important sector and OpenAI holds more compute than a token-constrained Anthropic. GPT-5.5 reviewed strongly while Opus 4.7 appears to be a bust, with users rolling back to 4.6. Enterprise and Codex are how OpenAI catches up.

   “I think Sam may end up being right here for the wrong reason, which is he missed on consumer, but enterprise is going gangbusters and is giving him the ability now, I think, to catch up on Codex.”

7. 07

   AI cyber models don't create vulnerabilities. They discover them.

   Sacks refuses the doomsday framing of Mythos and GPT-5.5 Cyber, the second model to complete a multi-step attack simulation end to end. The bugs were already in the code; a model that finds them first in defenders' hands hardens infrastructure. He expects every frontier model, including the Chinese ones, to have this capability within about six months, followed by a one-time upgrade cycle and then a new offense-defense equilibrium that feels normal.

   “it doesn't create the vulnerabilities. It just discovers them. The bugs were already in the code.”

8. 08

   All the software that runs the world gets rewritten in the next five or six years.

   Chamath sees cyber moving from humans exploiting humans to machines versus machines. Once that is the threat model, there is enough reason to rewrite the operational software underneath everything, increasingly by machines, and he expects the result to be more secure even as the threat rises.

   “Probably in the next five or six years, there'll be so much reason to rewrite all of the software that runs the world.”

9. 09

   Brockman's own diary says the real goal was to get Elon out.

   Jason reads from Greg Brockman's discovery diary in the Musk v. Altman trial: the stated aim was pushing Elon out, and if the company converted to a for-profit months later the public would correctly conclude they had not been honest with him. Friedberg cannot believe the diary exists at all. Elon is seeking $150 billion, a reversion to nonprofit, and the removal of Altman and Brockman; Polymarket sat at 42-43% that he wins even after the documents dropped.

   “The true answer is that we want Elon out. If three months later we're doing B Corp, then it was a lie.”

10. 10

    Vibe coding is a risk, not a tax. It will torch enterprise value and get people fired.

    Chamath's read on the Pocket OS incident, where an agent deleted a production database and its backups without confirmation, is that this was the preview, not the exception. Sacks calls it old-fashioned bugs at an edge case, but agrees AI still doesn't know what it doesn't know and needs supervision. Chamath goes further: somewhere a public-company employee tries to vibe-code their way out of a problem and destroys value at scale.

    “People will get fired because there will be some public companies where some goofball tries to vibe code their way out of something and they're going to torch the enterprise value.”

11. 11

    Retatrutide might be a de-aging drug, not just a weight-loss drug.

    Friedberg walks the phase-3 data on Lilly's triple agonist: the average patient dropped 37 pounds from 214 in 40 weeks versus 6 on placebo, with triglycerides down 41%, liver fat down 80% and A1C falling from 7.9% to 6%. The added glucagon agonist favors burning fat over muscle, which is why he thinks a low dose could cut weight while keeping muscle, and the anti-inflammatory effects point past obesity and diabetes. He expects FDA approval around mid-2027, with Lilly pricing it as the Mercedes above $50 tirzepatide.

    “reticulotide might actually be kind of a de-aging drug as well”

12. 12

    If states can override an EPA label, they can ignore the FDA and USDA too.

    Friedberg sat in on the Supreme Court's Monsanto/Roundup arguments, where the question is whether a federal FIFRA label preempts state failure-to-warn claims now that Chevron is gone. Bayer has paid $10 billion, reserved another $10 billion, and faces 90,000 open cases. His worry is the precedent: letting states interpret federal law around a regulator opens the door for every state to route around every federal agency.

    “if the states get to interpret federal law and ignore federal regulatory bodies, it opens up a whole new can of worms in terms of like all the states can start to ignore federal regulatory bodies, like the EPA or the FDA or the USDA”

My biggest takeaways from this week's All-In:

1. OpenAI's miss is a power problem, not a demand problem. Chamath reads the WSJ report on OpenAI missing its billion-weekly-user and revenue targets as a supply story: nobody in this market is short of demand, they are short of the electricity needed to generate tokens. Backlogs in power and components hurt OpenAI and Anthropic most and hand leverage to the hyperscalers, who end up trading compute for equity and control. His verdict on the misses: 'It is entirely 100 percent due to the supply of the power necessary to generate the output token.' [8:29]

2. There are no dark GPUs, so this is not Cisco in 2000. Sacks rejects the dark-fiber analogy Jason raises against roughly $725 billion of 2026 hyperscaler CapEx. In 2000 the fiber went dark and Cisco needed 25 years to regain its peak; today every GPU that comes online is consumed by demand for tokens, and that demand is what pulls the investment forward. He counts the simultaneous Google, Microsoft, Amazon and Meta beats as the moment the bull thesis got validated in a single afternoon. [46:55]

3. AI is now synonymous with the American economy, and it won't bring the job loss the doomers predict. Sacks puts AI at about 75% of GDP growth last quarter, with CapEx alone above 2% of GDP. That makes halting AI equivalent to halting the economy, which is why he thinks the political case against it collapses. On the labor side he sees a productivity boom rather than mass displacement, and treats 'eliminating all software developers' as the peak of inflated expectations that has already passed. [49:38]

4. In five years the hyperscalers look like bulky industrials. Follow the dollars to their suppliers. Chamath sees a structural shift from asset-light to asset-heavy capital markets. Around $700 billion a year of CapEx, collapsing free cash flow (Amazon down 97%) and heavy leverage turn Google, Microsoft, Amazon and Meta into bulky industrial businesses with no obvious valuation case at the end of it. His example of what they are willing to pay: Microsoft's Three Mile Island power agreement ran at more than twice the spot rate. The trade is to buy the underpriced asset-heavy suppliers the trillion dollars flows through. [43:30]

5. Any mature market ends up 4-2-1, and AI will be two of them: consumer and enterprise. Friedberg applies BCG's Rule of Three: a stable competitive market settles into a 4-to-2-to-1 share ratio among three players covering about 90% of the market. He splits AI into two such markets. Consumer trends toward a ChatGPT-versus-Google fight for first and second with Anthropic third; enterprise is led by Anthropic or Google. OpenAI's 900 million weekly users still dwarf Claude's, which he puts at sub-100 million. [13:02]

6. Sam will be right about his compute bet for the wrong reason: coding, not consumer. Sacks argues the $600 billion in compute commitments looked reckless against a consumer miss, but coding became the all-important sector and OpenAI holds more compute than a token-constrained Anthropic. GPT-5.5 reviewed strongly while Opus 4.7 appears to be a bust, with users rolling back to 4.6. Enterprise and Codex are how OpenAI catches up. [4:49]

7. AI cyber models don't create vulnerabilities. They discover them. Sacks refuses the doomsday framing of Mythos and GPT-5.5 Cyber, the second model to complete a multi-step attack simulation end to end. The bugs were already in the code; a model that finds them first in defenders' hands hardens infrastructure. He expects every frontier model, including the Chinese ones, to have this capability within about six months, followed by a one-time upgrade cycle and then a new offense-defense equilibrium that feels normal. [21:48]

8. All the software that runs the world gets rewritten in the next five or six years. Chamath sees cyber moving from humans exploiting humans to machines versus machines. Once that is the threat model, there is enough reason to rewrite the operational software underneath everything, increasingly by machines, and he expects the result to be more secure even as the threat rises. [25:04]

9. Brockman's own diary says the real goal was to get Elon out. Jason reads from Greg Brockman's discovery diary in the Musk v. Altman trial: the stated aim was pushing Elon out, and if the company converted to a for-profit months later the public would correctly conclude they had not been honest with him. Friedberg cannot believe the diary exists at all. Elon is seeking $150 billion, a reversion to nonprofit, and the removal of Altman and Brockman; Polymarket sat at 42-43% that he wins even after the documents dropped. [30:52]

10. Vibe coding is a risk, not a tax. It will torch enterprise value and get people fired. Chamath's read on the Pocket OS incident, where an agent deleted a production database and its backups without confirmation, is that this was the preview, not the exception. Sacks calls it old-fashioned bugs at an edge case, but agrees AI still doesn't know what it doesn't know and needs supervision. Chamath goes further: somewhere a public-company employee tries to vibe-code their way out of a problem and destroys value at scale. [57:45]

11. Retatrutide might be a de-aging drug, not just a weight-loss drug. Friedberg walks the phase-3 data on Lilly's triple agonist: the average patient dropped 37 pounds from 214 in 40 weeks versus 6 on placebo, with triglycerides down 41%, liver fat down 80% and A1C falling from 7.9% to 6%. The added glucagon agonist favors burning fat over muscle, which is why he thinks a low dose could cut weight while keeping muscle, and the anti-inflammatory effects point past obesity and diabetes. He expects FDA approval around mid-2027, with Lilly pricing it as the Mercedes above $50 tirzepatide. [1:00:59]

12. If states can override an EPA label, they can ignore the FDA and USDA too. Friedberg sat in on the Supreme Court's Monsanto/Roundup arguments, where the question is whether a federal FIFRA label preempts state failure-to-warn claims now that Chevron is gone. Bayer has paid $10 billion, reserved another $10 billion, and faces 90,000 open cases. His worry is the precedent: letting states interpret federal law around a regulator opens the door for every state to route around every federal agency. [1:13:22]

— All-In Podcast, "OpenAI Misses Targets, Codex vs Claude, Elon vs Sam Trial, Big Hyperscaler Beats, Peptide Craze" (2026-05-01)