AI Boom Archive
The euphoria, the warnings, and the bizarre — chronicled in real time. Hover a clipping to read it.
Jul 1, 2026
Meta's "excess compute" is a model demand problem, not a surplus
Rittenhouse Research @RHouseResearch
In March, Meta signed a $27B cloud services deal with Nebius. In April, Meta signed a $21B expansion of their cloud services agreement with CoreWeave. Two weeks ago, Bloomberg reported that Meta signed up to rent 1.6GW of data center capacity from Crusoe. And just 3 days ago, the Financial Times reported that Google had capped Meta's use of Gemini because of capacity constraints. And now all of a sudden, Meta actually has excess capacity? In reality, I suspect Meta looked at SpaceX's valuation and recent deals to sell excess capacity.. And came to the conclusion that investors will look favorably upon any willingness to monetize their capacity externally.. Which will give their share price some relief, and allow them to raise capital.. Which they'll use to buy and build more compute capacity, lol.
We are still massively short compute. $META and xAI are selling compute because there is no inference demand for their models. It's a compute allocation problem (too much compute in the hands of players with no internal use for it) not a compute surplus problem.
Jun 26, 2026
AI capex math doesn't add up
Arvind Krishna
So, let's ground this in today's costs because anything in the future is speculative. It takes about $80 billion to fill up a one-gigawatt data center. That's today's number. If one company is going to commit 20-30 gigawatts, that's $1.5 trillion of CapEx. To the point we just made, you've got to use it all in five years because at that point, you've got to throw it away and refill it. Then, if I look at the total commits in the world in this space, in chasing AGI, it seems to be like 100 gigawatts with these announcements. That's $8 trillion of CapEx. It's my view that there's no way you're going to get a return on that because $8 trillion of CapEx means you need roughly $800 billion of profit just to pay for the interest.
Excerpt from a longer interview on The Verge podcast; full conversation at link. Krishna also argues AGI from LLMs alone is impossible and would require subsequent breakthroughs (the LeCun view), possibly neurosymbolic AI.
Jun 26, 2026
Can LLMs come up with novel ideas?
Zhu Liang @paradite_
i'm really surprised that people don't see this.
It's mathematically true that llms can't come up with novel ideas, because the whole point of training is to reduce loss, gain rewards so that the model adhere to rules and ground truth.
if you have a model that can come up with novel ideas, it must have high loss during sft or rl.
"LLMs CAN'T COME UP WITH NEW IDEAS."
new ideas aren't out of distribution. they come from recombination, abstraction, analogy, and search.
the Wright brothers saw birds, bicycles, wings, engines, and then combined them into an airplane.
Boden's creativity framework is a useful starting point here. LLMs can do combinatorial and exploratory. Less clear on transformational.
Shevlin shares Margaret Boden's three types of creative thought (Boden, 2007, "Creativity in a Nutshell"): combinatorial (combining existing ideas into something new), exploratory (exploring possibilities within a domain), and transformational (radically new ideas that redefine a domain and its rules). Second link: https://x.com/dioscuri/status/2068625521567519179?s=20
Jun 26, 2026
Open-source won't dent frontier revenues
Simon Grimm @Simon__Grimm
I don't think open-source models will threaten frontier company revenues.
People often show this open-router statistics to suggest open-source models take up more market share. But the graph is misleading, as it doesn't show the fast-growing API use of frontier companies.
But even if distilled models take on easier, non-important tasks, this might not impact frontier company revenues much, as the most economically valuable and lucrative tasks are still done by frontier models.
In a new blog post, we explain why. simongrimm.substack.com/p/is-there-a-m…
Briefly, revenue concentration is explained by the most economically valuable tasks either only being doable by frontier frontier, or because relative ability is important (which is crucial across a wide range of tasks, e.g., pitching for a contract, or writing job applications). In this way, AI markets might resemble labor markets: while most hours/tokens are spent doing routine work at modest pay, the most valuable work commands most returns.
If commoditization eats into frontier company revenues, those doing the commoditization will also make no revenues, making it hard for fast-followers to spend anywhere near as much on R&D as frontier companies. So I think we should expect the lead of frontier companies to persist, especially as they clamp down on distillation.
Cheap models might provide an alternative for easy or unimportant tasks. But even that wouldn't change much about the revenue distribution between AI companies: once a task can be done by a commodity model, its price collapses. The revenue from this kind of work is small even if it creates lots of the value. In this world, the AI market would resemble something like the market for lawyers, where much work is routine and modestly paid, while a few superstars work on high-value problems and make enormous returns.
Pull-quote from the linked blog post (simongrimm.substack.com):
Jun 24, 2026
Hyperscaler FCF nosedive — warning or buildout?
Chamath Palihapitiya @chamath
At a very hi level, free cashflow = operating cash flow - capex.
The hyperscalers are deep into an investment cycle so they are consuming their operating cash flow. It is incorrect to look at this and assume their FCF has collapsed because operating cash flow has collapsed.
It hasn't. Capex has exploded.
This cycle should be eerily reminiscent of Amazon's approach over the past 20 years when they did the same thing related to e-commerce and AWS buildout. The question should be what moat did Amazon create at the end of that cycle and what kind of moat could the hyperscalers build now related to AI after this cycle?
This is a hell of a chart from @Nomura
Oil & Gas ramped up capex from 2010-2014 when oil was $110+ and the world was convinced we'd hit "Peak Oil" supply. Then Saudi started a price war with US Shale and those billions of dollars in capital were destroyed. These capex cycles happen all the time and never end well.
$AMZN Free Cash Flow actuals and estimates going out to 2030.
Note: link for the Hidden Small Caps Amazon chart (screenshot 3) not yet captured — add when available.
Jun 22, 2026
US AI capex dwarfs China's
Callum Williams @econcallum
Astonishing difference between US and Chinese AI capex, even after taking into account differences in PPP
Chinese hyperscalers (Alibaba, Tencent, ByteDance, Baidu): 8 (2022) · 18 (2023) · 36 (2024) · 57 (2025) · 102 (2026E) · 123 (2027E). ByteDance is not covered; GS extrapolates from comparable companies.
Jun 20, 2026
Automate 90%, then find new work for people
Dario Amodei @Overlap_Tech
"We are seeing right now that AI is making people more productive, but that's the usual hump. You automate 90% of the job. People are 10 times more productive in the other 10%. But eventually it gets close to 100%. The sequel to that is, well then you have to find something else for them to do."
Video clip posted by @Overlap_Tech ("Overlap: Business & Tech"); quote is Dario Amodei.
Jun 19, 2026
China's cheaper tokens erode US pricing power
Trevor Noren @trevornoren
FT: "Goldman Sachs analysts last month predicted that use of AI agents would result in a 24-fold increase in token consumption by 2030 and that the huge rise in demand would exacerbate a shortage of chips over the next 12 to 18 months. While token usage and AI spending by businesses continue to grow, efforts to curb costs could weigh on the growth of the world's largest AI labs such as Anthropic and OpenAI, which plan to go public later this year at near-trillion-dollar valuations. Since the start of the year, Chinese AI models have overtaken their US counterparts in token consumption, according to data from OpenRouter, an aggregation platform that allows users to access multiple AI models. China's cheaper energy and more efficient models have allowed the country's AI labs to charge less than leading US groups for tokens, giving China a new edge on the AI battleground."
Again, I believe market participants are underestimating the pricing power challenges US hyperscalers face, both due to domestic and international competition. As I wrote in my December report on "GenAI & Productivity" (sageroadresearch.com/collections/re…):
"While there's a lot of speculative fear about how a single LLM could rise to dominance and what that could mean for economic, societal, and political stability, we believe the bigger concern for investors today is how relative model parity could compromise pricing power. Tech giants have thrived on monopolies and duopolies for a decade or more. Now, they're in an LLM arms race where it's unclear when or even if ever leadership will be sustainable. We believe competition from akin models will apply downward pressure on pricing at least for the next three years."
Learn more about Sage Road Research here: sageroadresearch.com. Interested in subscribing? Message me.
FT link: ft.com/content/1d37cc…
Chart attached: "Chinese AI models have rapidly gained traction — Weekly usage in 2026 of top nine models on OpenRouter (trillion tokens)." China's line spikes to ~18T tokens/week by June 2026 vs. the US at ~5T.
Jun 16, 2026
Hyperscalers can't self-fund the capex boom
Epoch AI @EpochAIResearch
The end of the self-funded AI buildout?
Hyperscaler cash capex is growing much faster than cash inflows. On current trends, they will be unable to fully fund the AI infrastructure buildout with cash from operations by the end of this year.
Chart attached: "Hyperscaler capex is on trend to outpace their cash inflows by the end of 2026" (US $ billions, log scale). Operating cash flow is growing ~+23%/yr while cash capex grows ~+70%/yr; the two lines cross around Q3 2026. Quarterly sums across Microsoft, Amazon, Alphabet, Meta and Oracle, from SEC 10-K/10-Q filings. Source: Epoch AI (CC-BY).
Jun 3, 2026
AGI around 2030, foothills of the singularity
Demis Hassabis @deredleritt3r
"Ten years from now, I think we will realize that we were standing in the foothills of the singularity now... I believe that we're only a few years away from [AGI], maybe 2030 plus or minus a year...
I think [AGI] will be an enormous transformative technology, it's going to effectively be a new human era...
We can feel this year, I would say, even though I've been working towards this for 30 years, I think this year with the way the agents are working and tool use, it started to become really useful, still early days of it, but genuinely useful in people's workflows...
And it's not any one thing, it's several different technologies, several use cases, several things that I thought were maybe a bit further out, turned out to be now, that are coming together that make me feel that in aggregate.
I think society needs to hear that because we don't have long to prepare for what that means."
First "foothills of the singularity" mention — Google I/O, ~May 2026 — context: https://www.semafor.com/article/05/20/2026/google-exec-demis-hassabis-predicts-were-at-the-foothills-of-the-singularity
Apr 13, 2026
We'll soon accept AIs as conscious
Ray Kurzweil
Ray Kurzweil just said something that gave me pause.
He believes AIs will soon be so indistinguishable from conscious beings that we'll simply accept them as conscious — not because we'll have definitive proof, but because it will become useless not to.
He pointed out that people already have AI therapists, and some users are starting to treat them as genuinely conscious. As the technology improves, that acceptance will only grow.
Kurzweil thinks the shift won't take long: once AIs consistently show all the earmarks of consciousness, most people will just go along with it.
It's a quiet but profound prediction about how quickly our definition of "person" (or at least "mind") might change.
What do you think — how long until we treat AIs as conscious beings?
Relayed by @newstart_2024 (Camus) as a summary of Ray Kurzweil's remarks (not a verbatim Kurzweil quote), attached to a 1:06 "Moonshots" video clip. Per archive rules the video isn't stored — only this screenshot + the summary.
Mar 26, 2026
AI agents solve problems like evolution
Stephen Wolfram
Stephen Wolfram says the way AI agents solve problems may resemble biological evolution.
They try something.
See if it worked.
If not, they try again.That simple loop created every living thing on Earth.
It may also be how intelligence explores the computational universe.
Posted by @vitrupo, summarizing remarks by Stephen Wolfram, attached to a 1:53 video clip of him speaking. Per archive rules the video isn't stored — only this screenshot + the summary. Text is the poster's paraphrase, not a verbatim Wolfram quote.
Dec 1, 2025
Fully AI-run companies are 'far-fetched'
Satya Nadella
Satya Nadella doubts that fully AI-invented and AI-run companies are coming soon, calling that idea too far-fetched.
Satya Nadella instead points to "dark" infrastructure like data centers that run mostly autonomously after huge amounts of human work to build and configure them.
Satya Nadella expects stunning increases in automation over the next five to ten years, but still believes people will remain essential.
Satya Nadella emphasizes that even in heavily automated systems, humans design, oversee, and correct what the machines do.
Satya Nadella frames AI as a powerful extension of human capability rather than a total replacement for human agency.
Posted by @WesRoth, summarizing Satya Nadella's remarks (md meets: interview with Mathias Döpfner), attached to a 1:23 video clip. Per archive rules the video isn't stored — only this screenshot + the summary. Text is the poster's paraphrase, not a verbatim Nadella quote.
Aug 19, 2025
Seemingly Conscious AI: danger vs. real evidence
Mustafa Suleyman @mustafasuleyman
What I call Seemingly Conscious AI has been keeping me up at night - so let's talk about it. What it is, why I'm worried, why it matters, and why thinking about this can lead to a better vision for AI. One thing is clear: doing nothing isn't an option. 1/
Seemingly Conscious AI (SCAI) is the illusion that an AI is a conscious entity. It's not - but replicates markers of consciousness so convincingly it seems indistinguishable from you + I claiming we're conscious. It can already be built with today's tech. And it's dangerous. 2/
Why it matters: to be clear, there's zero evidence of AI consciousness today. But if people just perceive it as conscious, they will believe that perception as reality. Even if the consciousness itself is not real, the social impacts certainly are. 3/
[replying to Jason Crawford (@jasoncrawford): "isn't this just an absence-of-evidence vs. evidence-of-absence thing? or do you think there is positive evidence for AI consciousness?"]
I do, yes. especially looking beyond pure-text LLMs, AI systems have capacities and, crucially, computations that resemble those associated with, and potentially sufficient for, consciousness in humans and animals + evidence that, in general, computation is what matters for consc.
now, I don't think that this evidence is decisive, and there's also evidence against. but "zero evidence" is just way, way too strong. I think that AI's increasingly general capabilities and complexity alone is some meaningful evidence, albeit weak.
Counter: Robert Long (@rgblong), Aug 22, 2025 — https://x.com/rgblong/status/1958685038670717089?s=20. Both posts are excerpts of longer threads; full threads at the links.
Apr 3, 2025
AI 2027 intelligence-explosion forecast
Daniel Kokotajlo @DKokotajlo
"How, exactly, could AI take over by 2027?"
Introducing AI 2027: a deeply-researched scenario forecast I wrote alongside @slatestarcodex, @eli_lifland, and @thlarsen
Chart (ai-2027.com): an R&D-multiplier ramp through Superhuman Coder (4x AI R&D multiplier) → Superhuman Remote Worker (100x) → Artificial Superintelligence (2000x), timed 2027–2028. Marks a "branch point: slowdown or race?" and an "adversarial misalignment detected" point, with a forked path where "alignment gets solved" vs. "USG captured by AGI."
Mar 31, 2025
AI needs an impossible $2T by 2030
David Einhorn
McKinsey & Company estimates cumulative total spending of $6.7 trillion on data centers globally through 2030, including $5.2 trillion in capital expenditures on those equipped to handle AI processing. But where will the nearly $7 trillion come from? In 2024, before the major ramp-up in capital spending, the Magnificent 7 generated about $500 billion in combined cash flow. Roughly half went to buybacks and dividends, about a third to capital spending, and the small remainder was split among M&A, debt repayment, and cash accumulation. Even if they cut buybacks and dividends, the Magnificent 7 can't do the whole AI spend without taking on a lot of leverage.
The total book equity of the Magnificent 7 is about $1 trillion. Unspent dry powder in the private equity and venture capital industries stands at roughly $2.5 trillion and $300 billion, respectively. Wall Street's biggest year for IPOs and secondary offerings raised about $300 billion. Even if all this capital were dedicated to AI investment, it would still leave a multi-trillion-dollar gap – one that would need to be filled through debt and leases. For context, the record year for U.S. nonfinancial corporate debt issuance was $1.2 trillion.
And what about the revenue that might create a return on such a massive investment? Today, total global annual advertising spend is about $1 trillion, while all paid subscriptions in media, software, cloud, and related services accumulate to another $1 trillion.
According to Bain & Company, AI revenues would need to reach $2 trillion by 2030 to generate an adequate return on the AI investment. In other words, that would be about 100% of today's global advertising and subscriptions combined.
Something has to give.
Excerpt from Greenlight Capital's quarterly letter (David Einhorn). Date is approximate — inferred from the file name "Qlet2025-03"; adjust if you have the exact letter date.
Built Jul 01, 2026 · 19:50 · 16 entries · source of truth lives in entries/