AGI May Already Be Here This Year; What's Left Is Turning It Into Money
Chamath's judgment is that AGI has essentially existed since the start of this year, and the only difference is the pace at which closed labs choose to release it — a gap that will be closed within three to four months. The real bottleneck isn't benchmarks, it's how companies and ordinary people organise it into something with ROI.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AGI may already have arrived at the start of the year
Chamath's judgment is blunt: AGI has essentially existed since the start of this year, and the only difference is the pace at which closed frontier labs are agonising over releasing it. He expects those capabilities to be matched or nearly matched by a handful of closed and open alternatives within the next three to four months. So for most people the thing to watch isn't the benchmark chart but two things: top-tier intelligence is being widely obtained, and the cost per unit of incremental intelligence keeps being pushed down. What's left is the ‘messy middle’ — how companies and people organise it into something with ROI. His advice is keep calm and carry on.
— ChamathUse one bot to manage all your bots
He built a bot called improve my bots bot, had it walk through all his bots, fill in skills and instructions, and improve them once a day. The bot volunteered that merging two or three of them into one instruction set would be better, then executed all the changes itself; he estimates his bots got 30% to 40% better on average. Another cheap hack is to just ask your chief bot ‘what else should I be using you for’, and it will suggest the next step. The move here is to turn tuning into a continuously running daily task rather than a one-off manual chore.
— David SacksThis time it's real money, but 50x is still dangerous
Freeberg points out the key difference from the dot-com bubble: back then the whole boom was built on numbers that weren't dollars, like site hits and pageviews; what you see now is revenue, profit, and revenue and profit growth. Chamath adds that the Anthropic story, data centres, infrastructure, networking gear — all of it is real dollars moving, not speculative utilization. But he also issues a pricing warning: companies priced at 50 to 100 times top-line revenue is unsustainable — giving a one-on-one legendary founder like Elon or Travis 30 times revenue makes sense, but an unproven founder getting 50x or 100x is dangerous, because entry price matters. He also says that if this cohort of companies loses 95% of its market cap within six months, the long tail below will simply go under, with people carrying Aeron chairs out of the office.
— FreebergFounders should take money, but Series A cash-out is a negative signal
Chamath gives founders two pieces of advice: you have the right to sell 10% to 20% of your stake now, especially if this is your first company and you haven't yet built up a principal; and if you plan to delay a raise by six months so the valuation can double or triple, that's a huge mistake — if you can get $100 million, $250 million, $1 billion into the cap table at a solid valuation or 10 to 40 times revenue, take it. Companies sitting on hundreds of millions to billions in cash have optionality and can survive a real retreat. Freeberg explicitly objects to applying this to early-stage companies: a Series A founder wanting to cash out is, in his view, a huge negative signal; selling stock in a company with no revenue yet feels like a lack of confidence in your own company.
— ChamathTwo IPOs will release half a trillion in philanthropic money
Once Anthropic and OpenAI go public, the charitable giving or DAF money the two companies generate will together exceed half a trillion dollars, possibly more. That money will be allocated through one specific lens — effective altruism — and will inevitably shape outcomes, just as the capital of Soros, the Koch family and Sheldon Adelson each shaped outcomes. Chamath says he doesn't judge whether EA is right or wrong; the only problem is that the conflicts of interest haven't been adequately disclosed: A writes articles, A shares an office with B, B's wife works under someone at a frontier lab, B also holds $5 million of Anthropic, C works at Anthropic, everyone is roommates and co-invests. These relationships aren't a problem in themselves — the PayPal Mafia had its own version back then — the difference is not disclosing.
— ChamathThe main line of AI politics has shifted to open versus closed
For the past few years the main line of AI political debate was accelerationists versus doomers, and it cut across parties: there were progressives like Bernie Sanders who are anti-progress and anti-innovation, and conservatives like Steve Bannon who fear new technology. Now the debate is turning to open versus closed. Sachs's judgment is that the technology won't be stopped, because consumers are adopting it — data centres are being built because consumers and enterprises demand it, and adoption is pulling the whole thing forward. In the same segment Chamath rates the Hugging Face deal highly: creating, inside the largest and best-capitalised companies, a fortress against the madness of the closed-source oligopoly, letting the best ideas win in the market; he says he doesn't want to be rugpulled into an oligopoly or duopoly market structure by manufactured panic.
— SachsTwo hundred accounts are enough to fake a consensus
Axios reported that roughly 200 accounts suspected of coming from Chinese platforms tried to influence Americans against AI data centres. Jason draws a more general mechanism from it: a small number of social media accounts constantly bombarding and swarming the replies can shape opinion and make people wrongly believe a consensus exists. He distinguishes this clearly from real public opinion — the real consensus isn't that people dislike data centres, it's that a bunch of bots are telling influential people that people dislike data centres.
— JasonThe Pentagon takes 35% equity in the oil contract
North American Blue Energy Partners got 100-year concessions on 17 oilfields involving 65 billion barrels of reserves, reserves previously controlled by Russian, Chinese and Maduro entities. The US government controls 55% of the deal, of which the Pentagon takes 35% equity and the State Department gets the right to buy 20% of production at cost, with the agreement signed by Rubio and Hegseth with Venezuela's interim president. Faced with ‘should you seize another country's oilfields’, Sacks's answer is that this doesn't count as nation-building: no army, no government, no teaching American-style democracy, just a commercial transaction. His technical argument is that US shale produces light sweet crude, diesel and other distillates come from heavy crude, and Gulf refineries were originally designed for heavy oil — converting them to handle light oil is very expensive.
— Jason CalacanisIn their own words · checked verbatim
much much much three muches more capable models are coming soon. The next generation of models are going to be sobering for everyone.
Sam Altman (program quote)2:03
With the boom, it was all like metrics that weren't dollars. Yes. What we're seeing now is revenue and profits and growth in revenue and profits that we've never seen before.
Freeberg12:11
euphoria exists because markets are real. It's just that people are guessing how far forward to price the reality. And bubbles burst when you realize that you priced way too aggressively forward.
Chamath14:12
If you're a founder who's raising a series A and you're trying to take chips off the table, I consider that to be a huge negative signal.
Freeberg17:14
My single issue is I think that they don't disclose properly enough their obvious conflict of interest.
Chamath42:31
the consensus is not that people don't like data centers. It's that a bunch of bots are telling influential people that people don't like data centers.
Jason56:35
Well, I definitely don't think this is nation building. You know, we're not trying to stand up an army here like we did in Afghanistan. We're not trying to teach them about toxic masculinity. We're not trying to stand up a government or anything like that.
David Sacks1:21:05
And our refineries in the Gulf were built for heavy crude and it's very expensive to try and retrofit them to do the the light crude.
David Sacks1:23:09
Figures
| OpenAI's probability on Polymarket of having the best AI model by the end of 2026 | Soared from low single digits to over 20% | 7:07 |
| Instinct (personal AI assistant) valuation | $2.5 billion, still in private beta | 9:09 |
| San Francisco mansion prices | Currently $3,000 per square foot, selling at twice asking; the ultra-luxury market is expected to reach $5,000 per square foot | 15:12 |
| Anthropic IPO size comparison | All IPOs in San Francisco history combined are only a quarter of it | 15:12 |
| Combined philanthropic money after the Anthropic and OpenAI IPOs | Over half a trillion dollars | 41:29 |
| Accounts suspected of being from Chinese platforms influencing US opposition to AI data centres | About 200 | 56:35 |
| Venezuela oilfield concession term | 100 years, covering 17 oilfields | 1:20:03 |
| Reserves involved in the deal | 65 billion barrels | 1:20:03 |
| US government share of control in the deal | 55% | 1:20:03 |
| Pentagon equity | 35% | 1:20:03 |
Glossary
- effective altruism (EA)
- A philanthropic movement that began on Ivy League campuses, arguing for quantifying charitable impact and directing resources where the marginal return is greatest.
- p(doom)
- The EA community's estimate of the probability that AI causes human extinction; Chamath points to it as the community's new attachment issue.
- optionality
- The room to manoeuvre that cash on the balance sheet buys — when the market turns you still decide your own next step, rather than being forced to sell yourself.
How to listen
Founders, investors and engineers watching AI valuations and the bubble, the open-versus-closed fight, and US geopolitical energy deals.
If you only care about AI, the Venezuela oil segment after 01:20 can be skipped.