{"slug": "portents-of-doom", "title": "Portents Of Doom", "summary": "Elon Musk's SpaceX projected a $28.5 trillion total addressable market rivaling the U.S. economy, but analysts warn the AI bubble is deflating as subsidies end and enterprise demand wanes. The S&P 500's cyclically-adjusted P/E ratio, corrected for an earnings bubble and accounting windfalls, may reach 75.7x—far exceeding historical extremes and signaling an unprecedented valuation bubble.", "body_md": "Elon Musk is the world champion of totally implausible projections, and Kim Khan reported on a personal best in *SpaceX sees total addressable market rivaling size of the U.S. economy*:\n\nThe $28.5T forecast compares to U.S. Q1 2026 nominal GDP of nearly $32T, with the estimate for the market of AI enterprise applications of $22.7T about 70% of total U.S. economic output.\n\nSam Altman and Dario Amodei just aren't this good, but their projections of their Total Available Market (TAM) are still turning out to be vastly optimistic. In *AI's Affordability Crisis* I showed evidence that the AI platforms could no longer afford the massive subsidies they were using to artifically inflate demand for their product, and that reducing the subsidies had made their enterprise customers reconsider their enthusiasm for deploying them. This is leading to investors belatedly realizing that AI platforms' projections of their TAM and thus their valuations are totally implausible.\n\nThis re-calibration is just one of the many signs that the AI bubble is about to deflate. Below the fold I present a necessarily incomplete list of them, which I will try to update as more appear.\n\n-\nThe AI bubble isn't just an equity bubble, as Torsten Slok explains in\n*AI Is Penetrating Every Corner of Financial Markets*:\n\nAI now accounts for nearly half of all IG issuance, 87% of VC funding and a growing share of HY, underscoring how deeply the AI investment cycle has penetrated every corner of finance.\n\n- Bryce Elder's\n*This is nuts upon nuts. When’s the crash?* compares this bubble with its two biggest predecessors:\n\nHere’s the title page of this month’s Panmure Liberum market update from strategists Joachim Klement and Francisca Reis. Our emphasis in bold below:\n\nIn 1929, the cyclically-adjusted P/E-ratio (CAPE) of the S&P 500 reached 32.6x according to Prof. Robert Shiller’s data. This was 1.8 standard deviations above trend at the time. In 2000, the CAPE reached 44.2x, or 3.3 standard deviations above trend – a clear sign of a bubble. However, as our chart below shows, earnings in both instances were within normal range, less than one standard deviation above trend.\n\nToday, the CAPE is at 41.0x, or 2.9 standard deviations above trend. Once again, we are clearly in bubble territory for stock market valuations. However, unlike in previous bubbles, we are having extremely high CAPE at a time when earnings themselves are 1.8 standard deviations above trend. In other words, we are in a valuation bubble at a time when earnings are in a bubble themselves.\n\n**If we correct for the earnings bubble, the current CAPE would be 67.6x or 4.6 standard deviations above trend, a bubble that surpasses anything ever seen in US history by an extreme margin. If valuations followed a normal distribution (which they don’t, so don’t take this literally), this would happen in 0.00019% of months or once every 43,432 years.**\n\n- But Panmure Liberum is underestimating how bad things are. Baolian Wang's\n*The $69 Billion Mirage: How an Accounting Rule Inflated S&P 500's Q1 Earnings by 12%* explains:\n\nA massive chunk of this quarter’s blockbuster “growth” didn’t come from selling more software, shipping more microchips, or delivering more packages. Instead, it came from an accounting rule that forced massive, illiquid “paper gains” onto the income statements of tech giants.\n\nIn Q1 2026 alone, just three companies—Alphabet, Amazon, and Nvidia—reported a staggering $69.2 billion in non-operating windfall under their Other Income and Expenses (OI&E) lines. When you run the macro numbers, this single accounting phenomenon artificially inflated the entire S&P 500’s quarterly earnings by about 12%.\n\nThat 12% takes the factor from 67.6 to 75.7. Wang notes that correcting for the 12%, \"the Q1 2026 earnings growth rate will not be very different from the 5-year average of 16%.\" In other words, the bubble is feeding upon itself — increased stock prices causes increased earnings causes increased stock price ... But suppose, for example, that OpenAI were to suffer a down round. Then Alphabet, Amazon, Nvidia and others who included paper gains in the \"other income\" on the way up would have to include paper losses in their income on the way down, amplifying the crash. OpenAI's last round valued the company around $750B and they were planning an IPO for at least $1T, but had to postpone it.\n- Back in March Jared James Grogan published\n*The End of the Foundation Model Era: Open-Weight Models, Sovereign AI, and Inference as Infrastructure* setting out the forces changing the structure of the AI market:\n\nThe foundation model era — roughly 2020 to 2025 — is over. The forces that defined it have inverted. Open source models have reached frontier performance while inference costs approach zero, exposing what was always structurally true: pre-training large language models at scale is not a durable competitive moat. The US government's formal designation of Anthropic as a supply chain risk in February 2026 accelerated a transition already underway — but did not cause it. The paper argues that the AI industry is restructuring simultaneously along four axes: economic, as the circular financing structure that inflated foundation model valuations collapses; technical, as the pre-training scaling paradigm gives way to post-training optimization, test-time compute, and agentic composition; commercial, as application-layer integrators displace the foundation model companies whose commodity they now consume; and political, as the government asserts its historic role as gatekeeper of strategic technology. These are not separate disruptions. They are one structural shift, arriving together.\n\nGrogan further [argues that](https://arxiv.org/pdf/2604.06217v1):\n\nThe most consequential and least-discussed dimension is the permanent divergence between commercial AI and a classified national security AI track — built on different data, governed by different rules, and developing capabilities the public ecosystem cannot see, measure, or govern. Like every dual-use technology that has altered the calculus of state power, AI is being brought under government authority not by design but by the structural logic of what it is. The paper further argues that open-weight models are the counterintuitive instrument of sovereign control: a government that holds the weights commands the capability on its own terms, without dependence on vendor policy, financial continuity, or personnel clearance. The apparent openness of distributed model weights is, from a deploying government's sovereignty standpoint, the most governable architecture — because what cannot be withdrawn by a vendor's API policy cannot be taken away.\n\nGrogran implicitly assumes that AI is useful but that the current margins and thus the valuations aren't sustainable.\n- Carl Franzen's\n*DeepSeek-V4 arrives with near state-of-the-art intelligence at 1/6th the cost of Opus 4.7, GPT-5.5* explains the impact of open-weights models on pricing:\n\nDeepSeek-V4-Pro is priced through its API at **$1.74 USD per 1 million input tokens on a cache miss and $3.48 per million output tokens.**\n\nThat puts a simple one-million-input, one-million-output comparison at **$5.22**. With cached input, the input price drops to **$0.145 per million tokens**, bringing that same blended comparison down to **$3.625**.\n\nThat is dramatically cheaper than the current premium pricing from OpenAI and Anthropic. GPT-5.5 is priced at **$5.00 per million input tokens** and **$30.00 per million output tokens**, for a combined **$35.00** in the same simple comparison.\n\nClaude Opus 4.7 is priced at **$5.00 input** and **$25.00 output**, for a combined **$30.00**.\n\nSix times cheaper for an equivalent product is likely to cause pricing pressure on the incumbents, who need to raise not reduce prices. DeepSeek is likely also subsidizing usage, but they do have real advantages. First, they have fewer resources so are forced to to be inventive. Second, the 40% of infrastructure capex that isn't the racks is much cheaper in China. Third, the power component of opex is much cheaper and more available in China. The [result is](https://venturebeat.com/technology/deepseek-v4-arrives-with-near-state-of-the-art-intelligence-at-1-6th-the-cost-of-opus-4-7-gpt-5-5):\n\nIn practical terms, DeepSeek does not need to win every leaderboard row to matter. If it can deliver near-frontier performance on many enterprise-relevant agent and reasoning tasks at roughly one-sixth to one-seventh the standard API cost of GPT-5.5 or Claude Opus 4.7, it still forces a major rethink of the economics of advanced AI deployment.\n\nDeepSeek-V4-Pro-Max is clearly the strongest open-weight model in the field right now, and it is unusually close to frontier closed systems on several practical benchmarks.\n\nWhile GPT-5.5 and Claude Opus 4.7 still retain the lead in most direct head-to-head comparisons across the company's benchmark charts, DeepSeek V4 Pro gets close while being dramatically cheaper and openly available.\n\n-\nDan Davies agrees that the margins aren't sustainable but differs as to why in\n*tokenalysis and john henry*:\n\nWhich then brings to mind another issue – how confident are we in the pricing power that underpins that 60% gross margin in the first place? In the last paragraph I was talking about the R&D equivalent of a price war, but the normal kind is also possible. The combination of price-sensitive B2B customers, big fixed costs and rewards going to the dominant player doesn’t suggest to me that pricing power is going to be sustainable indefinitely.\n\nBut, I think there’s a danger of missing the big picture here. Which is that, when large companies are telling their employees to be sensible and use AI tokens wisely, then the game is up. The race is over and [John Henry won against the steam hammer](https://en.wikipedia.org/wiki/John_Henry_(folklore)). If you need a human being in the loop to decide on the allocation of AI tokens, then all those predictions of mass redundancy are gone.\n\n-\nIn\n*A Slower AI Payoff Would Be Everyone's Problem*, Torsten Slok is mainly focused on the fallout we can expect even if the bubble deflates gently. He asks\n\n**But what if the payoff takes longer than consensus assumes?** That question is particularly pressing given that token prices continue to decline and Chinese models are gaining ground, both in their share of the world's most-used models and in token usage, where they now lead their US counterparts among the top 20 models,\n\nSlok provides two charts, the first tracks market share by country of origin monthly since January 2025 among the top 50 models. It is bad for the US, showing a steady erosion of US market share until, eyeballing it, as of May 2026 it is roughly a 60/40 US/Chinese split.\n\nClearly, the market is voting with its feet that the prices charged by US AI companies are unsustainable.\n\nThe second compares monthly token use among the top 20 models by country of origin between May and June this year. It is much worse for the US.In May the Chinese had 80% of the US usage. In June they had 185% of the US usage. If this rate of market erosion were to continue for a few months it would be impossible for investors to continue to imagine the golden future awaiting OpenAI, Anthropic, xAI, Meta, Oracle and the neoclouds.\n\nSlok's analysis of the fallout of the bubble deflating is worth reading. This is an issue I plan to return to in a future post.\n-\nKevin Walmsley's\n*We were wrong about DeepSeek. Now Chinese AI companies export trillions of AI tokens* also focuses on the erosion of US market share:\n\nChinese AI providers are making more money; Zhipu’s revenues went up 60 times in Q1 compared to last year; Alibaba is the company behind Qwen, and their revenues are up 15 times just since the beginning of this year. The lion’s share of the profits, though, are still being realized by the hardware side; the chipmakers, at least for now. The model companies’ revenues are rising, and steeply. But so is their cost of compute.\n\nThat dynamic is causing some Chinese labs to raise prices; the cost to use Tencent Cloud increased by over five times in March, and Alibaba, Zhipu, and ByteDance also hiked prices.\n\nDeepSeek, however, went the other way. Their latest version is priced at just a fourth of their introductory product, and they rolled out dynamic pricing that is aimed at corporate users, who use their models during the workday.\n\nThis price difference is having an [:](https://kdwalmsley.substack.com/p/we-were-wrong-about-deepseek-now)\n\n[\nBut even with these price increases, ](https://kdwalmsley.substack.com/p/we-were-wrong-about-deepseek-now)[Chinese models just cost far less](https://www.businesstimes.com.sg/companies-markets/telcos-media-tech/ais-token-economy-revolution-creates-new-china-tech-winners) than those on offer from Silicon Valley, and explains those big jumps in the exports, we can say, of Chinese AI tokens. The LLM’s out of China are [“90% as good at 10% of the cost”](https://www.businesstimes.com.sg/startups-tech/technology/cheaper-ai-better-soaring-bills-are-reshaping-how-businesses-choose-models), and American firms are buying more tokens from Chinese companies. In early 2025, token demand from Chinese LLM’s was about zero. Even the release of DeepSeek didn’t move the needle much, but by the end of the year the secret was out, and it’s been a [choppy but steady ride up to 46%](https://finance.yahoo.com/technology/ai/articles/chinese-ai-models-now-capture-020440715.html), today.\n\nWhy would you invest in a company whose competitors were “90% as good at 10% of the cost” and which was rapidly losing market share?\n-\nIf the margins, and thus the rational valuations, of AI companies are unsustainable, how long can the market remain irrational? In\n*The Second Derivative: Why No One Understands the AI Boom* Groundbreaker starts by examining the 2008 crash:\n\nThe implicit underwriting assumption, shared by originator and borrower alike, was that the loan would never reach its reset: rising home values would manufacture equity, the borrower would refinance into a fresh teaser and the clock would start again. **The structure was a treadmill, and the treadmill was powered by appreciation**. It worked spectacularly while it worked. Nearly four in five subprime hybrid ARMs originated in 2003 had been refinanced away by the end of 2006.\n\nNow watch the timing. National home-price appreciation did not crash in 2006. It *decelerated*. The year-over-year rate of gain, which had run in the mid-to-high teens through 2004 and into early 2005, began bleeding off - still positive, still printing green, but slowing. **Prices were higher than they had ever been. And yet, with prices at their peak and still rising, subprime delinquencies inflected upward**.\n...\n\nThe deceleration was endogenous to the structure; the structure required ever-accelerating prices to keep refinancing its way out of its own reset schedule, and no series accelerates forever.\n\nThe second derivative was always going to roll over. When it did, the first derivative followed it down through zero, negative equity spread from the margin inward, and **the defaults the market insisted were caused by “falling prices” had in fact begun a year earlier, when prices were still rising but had stopped rising ***faster*.\n\nWhy does the 2008 analogy [apply this time](https://www.groundbrkr.com/p/the-second-derivative-why-no-one?hide_intro_popup=true)?:\n\n**The market is pricing AI as a *** technology* cycle when its actual anatomy is that of a *credit-driven real estate* cycle - which is precisely why the 2008 mechanics apply - and the two break for entirely different reasons.\n\n...\n\nWalk down the AI build-out and every feature is a property development in disguise: a data center on entitled land, financed with debt against the structure and leased to tenants on take-or-pay terms. **This is not a software business that happens to own servers. It is a real estate business that happens to compute**.\n\nThe AI build-out is being [financed with debt](https://www.groundbrkr.com/p/the-second-derivative-why-no-one?hide_intro_popup=true):\n\nWhen a hyperscaler or a neocloud reports a record capex figure, the financial press reads it as confidence, as proof of demand. Read it instead as origination volume. Each gigawatt of committed build is a loan extended to whichever tenant has signed the take-or-pay beneath it, and the credit quality of that loan is precisely the credit quality of the tenant. The market is celebrating loan growth and calling it revenue growth.\n\nLoans in this market take the form of Remaining Performance Obligations (RPOs). The biggest borrower in this market is [OpenAI](https://www.groundbrkr.com/p/the-second-derivative-why-no-one?hide_intro_popup=true):\n\nOpenAI has committed to pay for compute on a scale without precedent in corporate history: multi-year, take-or-pay capacity contracts whose aggregate obligations run to the hundreds of billions of dollars. Against them sits an operating business that does not yet earn a profit - revenue real, large, and growing quickly, but short of covering the company’s own cash burn and nowhere near covering the contracted payments.\n\n**Those payments are therefore not serviced out of earnings. They are serviced out of financing, and financing, for a borrower in this position, is available on a single condition: that each new round price above the last.**\n\nDoes OpenAI's financing meet [this requirement](https://www.groundbrkr.com/p/the-second-derivative-why-no-one?hide_intro_popup=true)?:\n\nMeasured as a level, OpenAI’s valuation is the most remarkable appreciation in the history of private markets - roughly $86 billion in early 2024, then about $157 billion, $300 billion, $500 billion, and approximately $852 billion by the spring of 2026. Measured as a rate of change, the same series inverts: the round-over-round step-up ran 1.83×, 1.91×, 1.67×, 1.70×, and falls to roughly 1.23× implied by the reported public-offering target. Private marks are inherently lumpy - negotiated, episodic, set by a handful of insiders - so no single step is decisive. But the trend is unmistakable: it bends down, and it bends hardest at the one mark set by the deepest, most unforgiving pool of capital - the public market. The implied IPO step-up is both the lowest in the sequence and the hardest to negotiate, and it is the one the structure must actually clear. This arithmetic is also the most probable explanation for OpenAI’s recent IPO delay.\n\nIn the same way that the private stocks generating unrealized gains are not liquid assets, neither are the [Remaining Perfoemance Obligations](https://www.groundbrkr.com/p/the-second-derivative-why-no-one?hide_intro_popup=true):\n\nAn RPO is not a liquid asset; it is a forward contractual commitment - a promise of future payment in exchange for future compute. And a multi-year commitment is worth exactly the creditworthiness of the entity on the other end of it. When that entity is investment-grade and cash-generative, the backlog is what it claims to be: high-quality visibility, merely deferred. When that entity is a pre-profit company that loses tens of billions a year and can pay only by continuously refinancing its own equity valuation, the backlog is something else entirely. It is a subprime commitment, used to justify massive, un-depreciated capital expenditure, reported to shareholders as structural strength.\n\nNow price the credit quality of that book. Of roughly $2.1 trillion in aggregate contracted backlog across the four big platforms, about half - on the order of $1.05 trillion - is owed by OpenAI and Anthropic. Microsoft’s book is about 49% these two names; Oracle’s is 54%, with roughly $300 billion owed by OpenAI alone; Google’s is 43%; Amazon’s is 51%.\n\nThe hyperscaler has, in economic substance, extended a concentrated, unsecured loan to cash-burning tenants. The RPO that Wall Street values as forward revenue is, in reality, a credit exposure to borrowers with no operating income.\n\n-\nMajor financial institutions are similarly concerned. In their\n*2026 Annual Economic Report*, the Bank for International Settlements writes:\n\nIn the near term, the ongoing AI investment boom raises questions about the sustainability of the current economic expansion. The five largest hyperscalers are set to spend over a trillion US dollars on AI-related capital expenditure from 2025 through 2026. These commitments are outpacing earnings and the free cash flow of these firms, leading some to issue debt to raise additional financing (Graph 11.A). This investment race may be partly driven by the perception that only a small number of players with superior technology will ultimately dominate the market shares. The intense competition raises the risk of firms over-committing resources to investment projects with still uncertain returns, leaving all firms vulnerable to disappointments in AI payoffs. Model analysis based on such contest motives highlights the downside risk of current AI exuberance. As competitive pressure drives capex higher, the net economic surplus – the total payoff less investment costs – declines for the sector as a whole and could turn negative in adverse scenarios (Graph 11.B). Disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted investment bust, with potential knock-on effects on financial conditions (see below).\n\nThe BIS has noticed another [looming problem](https://www.bis.org/publ/arpdf/ar2026e.pdf):\n\nAnother risk is that the AI boom runs into a supply side roadblock. The AI build-out has recently been facing growing bottlenecks in electricity, advanced semiconductors and grid equipment. Fast-growing demand for computing power is already pressuring electricity prices and input costs, with potential spillovers to inflation. Looking ahead, these temporary shortages may also amplify over-investment, as firms attempt to lock in future capacity through long-dated contracts that further expose them to any disappointments in demand.\n\nUnlike Groundbreaker, the BIS uses more conventional [historical analogies](https://www.bis.org/publ/arpdf/ar2026e.pdf):\n\nHistorical episodes of investment booms offer instructive parallels (Graph 11.C). The canal mania of the 1830s, the British railway mania in the 1840s, the electrification exuberance of the late 1920s (roaring 20s) and the dotcom boom of the late 90s all shared one common trait: a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify. These episodes ended with an eventual reversal in investment, inducing economy-wide recessions. The scale and pace of the current AI investment boom accompanied by expectations of large productivity payoffs bear resemblance to these precedents, highlighting potential downside risks in the near term.\n\n-\nAnd Eric Katz reports that the US\n*Treasury Has an Internal Report Warning About the Dangers of an AI Bubble*:\n\nA draft report inside the Treasury Department is set to warn of the risks posed by the artificial intelligence market, likening key aspects of it to the dotcom bubble that upended the U.S. economy when it burst in the early 2000s.\n\nThe document, the existence and contents of which have not been previously reported but was obtained by NOTUS, is a significant departure from the Trump administration’s public tone, which has focused on encouraging unrelenting investment to unlock exponential growth.\n\nCareer Treasury analysts found that AI firms are more deeply entrenched in the U.S. economy than their dotcom predecessors and pose significant risk to the entire system if financial conditions change, productivity goals are missed or various choke points stymie growth.\n\n-\nVanderbilt University's Asad Ramzanali has a detailed look at the range of impacts from the burst bubble, and an set of optimistic suggestions for policy responses in\n*After the AI Crash*. He frames the problem thus:\n\nCompanies are investing trillions of dollars based on tens of billions of dollars in revenues. Analysts at J.P. Morgan anticipate $5 trillion of AI infrastructure investment in the next five years. They estimate that the industry will need to generate annual revenues of $650 billion to justify this level of investment, while consultants at Bain & Co. estimate $2 trillion in needed annual revenues. Yet, OpenAI and Anthropic earned $13 billion and $4 billion, respectively, in 2025 revenues. OpenAI’s own financial expectations suggest negative cash flow until 2030, and Anthropic expects a small profit no sooner than 2029. Alphabet, Meta, Amazon, and others may experience increased marginal revenue from integrating AI into existing products, but that is far from certain.\n\nNote that over the next 5 years around $3T (60% of $5T) of the investment in AI infrastructure goes in to buying the hardware which should be [fully depreciated](https://blog.dshr.org/2025/10/depreciation.html) over much [less than 5 years](https://blog.dshr.org/2026/02/mind-gaap-again.html). So the gap between the investments and the revenue is much bigger than it appears.\n\n-\nThe cracks are starting to show. Reuters reports that\n*Blackstone's QTS terminates Digital Gateway data center project in Virginia*:\n\nBlackstone's QTS said on Thursday it had terminated its planned Digital Gateway data center project in Virginia and withdrawn the associated filings after years of planning and regulatory review.\n\nThe data center operator has faced years of local opposition and litigation over the project, despite it being approved by the Prince William Board of County Supervisors.\n\n-\nAmong the hyperscalers, Oracle is the most exposed because, as\n[Ed Zitron noted](https://www.wheresyoured.at/the-ai-industry-is-losing/):\n\nAnd Oracle ... is a company that, even before the AI bubble, was massively indebted. It just so happens that, as a result of its tryst with OpenAI, Larry Ellison saw fit to twist the debt knob to eleven.\n\nNow, Brody Ford reports that *Oracle Warns Its Splurge on AI Data Centers May Not Pay Off*:\n\nSix firms alone — including Oracle, Microsoft and Meta — [have committed $850 billion](https://www.bloomberg.com/news/articles/2026-06-24/meta-microsoft-lead-850-billion-boom-in-data-center-leases) for data centers leases that haven’t begun yet. Oracle holds the largest share of these commitments owing to its $300 billion Stargate contract with OpenAI.\n\nWhen Oracle mentions the risk of nonpayment, the unnamed elephant in the room is OpenAI. As part of the [Stargate deal](https://www.bloomberg.com/news/articles/2025-01-21/trump-to-announce-joint-openai-softbank-oracle-ai-investment) with the AI company, Oracle is developing massive data centers across the country to provide cloud computing power. For this plan to work, OpenAI needs to pay its Oracle Cloud Infrastructure bills.\n\n“Some of our customers may be highly leveraged and subject to their own operating and regulatory risks and, even if our credit review and analysis mechanisms work properly, we may experience risks of non-payment and non-performance in our dealings with such parties,” Oracle said in the filing.\n\n-\nA month ago Laura Benitez\n*et al* reported that *SoftBank Attempt to Get $6 Billion OpenAI Margin Loan Stalls*:\n\nSoftBank Group Corp.’s talks with potential creditors to raise at least $6 billion from a margin loan backed by its OpenAI stake have stalled, people familiar with the matter said, just weeks after the Japanese conglomerate cut its initial target from $10 billion.\n\nBut now Echo Wang reports that *SoftBank renews talks for $10 billion loan against OpenAI stake, adds concessions, sources say*:\n\nSoftBank Group has reopened talks with a consortium of lenders for a $10 billion loan backed by its stake in OpenAI, after earlier attempts to secure a loan stalled over concerns about the difficulty of valuing private companies, two people familiar with the matter said.\n\nTo make lenders more comfortable, the Japanese technology investor is offering to guarantee repayment of the loan, giving banks recourse to SoftBank if the OpenAI shares pledged as collateral lose value, the people said.\n\nThis all seems to indicate that potential lenders, such as banks, are highly skeptical of the value of OpenAI stock.\n-\nDespite Grok being so bad that employees use Claude, Musk is touting SpaceX as an AI company. This resulted in the most overvalued IPO in history, which failed to raise enough money to avoid the immediate need to borrow $25B. Nir Kaissar's\n*SpaceX Is Junk. That’s What the Bond Market Says* reports on the bond market's reaction:\n\nRatings companies and the bond market have very different views about how things are going. SpaceX’s bonds have an average rating of BBB across the three majors, Moody’s Ratings, S&P Global Ratings and Fitch Ratings, according to credit scores compiled by Bloomberg. In the alphabet soup of bond ratings, it’s the lowest grade still considered quality before falling into junk territory.\n\nThe bond market has other ideas. There, quality is judged by a bond’s credit spread or the additional yield it offers above Treasuries with similar maturity. The wider the spread, the lower the quality. Corporate bonds with a BBB rating are trading at an average credit spread of 0.92 percentage point. SpaceX’s bonds, by contrast, trade at a significantly greater average spread of 1.62 percentage points across maturities, higher than BB rated junk bonds’ average spread of 1.55 percentage points.\n\nThe bond market seems to agree with Softbank's lenders about the AI bubble.\n-\nOpenAI and Anthropic are competing for the next trillion-dollar IPO. Both would need to distract investors from their massive losses by focusing on growth. Recently, Anthorpic has been growing faster than OpenAI, so Keach Hagey and Berber Jin report that\n*OpenAI Considers Drastic Price Cuts, Anticipating War for Users With Anthropic*:\n\nOpenAI is considering drastically lowering the prices it charges users as it seeks to win customers from its rival Anthropic.\n\nThe company is weighing significant cuts to what it charges for tokens, the unit of measurement artificial-intelligence firms use to bill for their products, according to people familiar with the matter. The move would be in anticipation of similar cuts the company expects at Anthropic, the people said.\n\nNow Kurt Wagner reports that *Zuckerberg Pledges ‘Aggressive’ Pricing With Meta’s First Pay-to-Use AI*:\n\nMeta will also introduce a new Meta Model API system, which will be used to collect fees from developers. Its API pricing is roughly 25% of the cost advertised by other top models from OpenAI and Anthropic PBC. Developers will be able to use Meta’s model for free, but only up to a point; they’ll be required to pay for access after reaching a certain token threshold, Zuckerberg said.\n\n“The pricing from some of the other labs is very extreme and has very high margins,” Zuckerberg said, underscoring that his strategy is to get Meta’s technology in front of as many people as possible. “We think that there’s a real ability to be able to offer frontier or very high-level intelligence at a much more affordable cost.”\n\nAggressive pricing means Meta Model API will still be 50% more expensive than DeepSeek but not 50% better. Planning to reduce current income in the lead-up to an IPO is an unusual move, but it is a response to *AI's Affordability Crisis*.\n-\nData center demands for power are having serious impacts elsewhere in the economy, as Jeremy Hsu reports in\n*Data centers’ energy demand threatens Trump’s “Made in America” plan*:\n\nFactory electricity bills are generally rising faster than those for other business customers or residential customers, [according to a Reuters analysis](https://www.reuters.com/business/energy/big-tech-data-centers-are-driving-up-power-bills-americas-rust-belt-factories-2026-07-07/). It highlighted the example of the Belden Brick Company, a 141-year-old brick manufacturer in Ohio, whose electricity bills have soared from $1,600 to $12,000 per month due to a higher monthly capacity charge in the 13-state region served by the grid operator PJM Interconnection.\n\n...\n\nThe Ohio-based steelmaker Metallus described its electricity costs as having jumped by 70 percent since 2024, leading the company to pay an extra $15 million in energy costs annually.\n\nThe higher electricity costs for manufacturers coincide with the attraction of large AI data center projects with [substantial electricity needs](https://arstechnica.com/ai/2026/07/googles-ai-buildout-drove-37-increase-in-electricity-use-in-2025/) to many states in PJM territory. That data center growth has driven up PJM’s capacity prices—paid to power generators according to supply-and-demand forecasts—from $28.92 per megawatt-day in 2024 to $329.17 per megawatt-day in 2026, according to Reuters’ reporting.\n\n-\n|\n| Source |\n\nAs I see it, there are three separate markets for LLMs. First, there is an *embedded* market that runs on low-cost, low-power hardware and open-weights models. *Small AI Models Gain Traction Around the World* by David Berreby provides examples:\n\nFor example, a [drone-based system developed by Bala Murugan and colleagues](https://www.science.org/doi/epdf/10.1126/science.adw7713) at the Vellore Institute of Technology, in India, takes photos of cashew plants and quickly identifies those with splotches that indicate disease. All the processing takes place on the drone itself, so there’s no need for a computer on-site, nor for a connection to a central server.\n\nUsing small language models trained for a specific problem, and sometimes running on cheap, low-power devices, other small-AI implementations have been developed to identify [ant infestations in a Uruguayan vineyard](https://universe.roboflow.com/juan-abedala/deteccion-hormigas-cortadoras), [detect the presence of malaria-carrying mosquitoes in a number of nations](https://dl.acm.org/doi/fullHtml/10.1145/3524458.3547258), and [run electrocardiograms from an Arduino device in parts of Brazil](https://link.springer.com/chapter/10.1007/978-3-031-49407-9_63) that lack access to more complex equipment.\n\nIt isn't just that small hardware, such as the Raspberry Pi 5, is getting more powerful but [also](https://spectrum.ieee.org/small-language-models-ai-pharmaceuticals):\n\nthe shrinking footprint of language models. Both Google DeepMind’s [Gemma 4](https://deepmind.google/models/gemma/gemma-4/) (released in April) and Alibaba’s [Qwen 3.5](https://qwen.ai/blog?id=qwen3.5) are “fantastic” for small AI, Rovai says. Both models are “open weight,” meaning users can adjust the connections between parameters to suit their needs. This makes it easy, for example, “to take a lot of data from, say, the milk industry and retrain the model specifically on that,” Rovai says.\n\nThe hyperscalers and AI platforms like OpenAI and Anthropic will garner no income from this market, [because](https://spectrum.ieee.org/small-language-models-ai-pharmaceuticals):\n\n“I think the future of AI is not like one giant model, at a center. I think it’s millions of small, precise models deployed at the edge, each one solving like a specific problem, a specific context,” Alonge says. This is partly because much of humanity—including people in parts of rich countries as well as the developing world—lives without access to cutting-edge frontier models. But, he says, it’s also because those models are not sustainable.\n\n“If someone is not subsidizing it, most people will not be able to afford those models. So those of us who are said to be small-AI developers are the ones who will have to build for the majority of the world,” Alonge says.\n\n-\nSecond, there is a\n*consumer* market. Last month, Nvidia announced a product for the prosumer laptop market with significant LLM capabilities. @pramodchandrayan described it in *NVIDIA Just Put a 120-Billion-Parameter AI Model in Your Laptop. Here’s What That Actually Changes.*:\n\nAt Computex 2026 in Taipei on June 1st, CEO Jensen Huang announced the RTX Spark superchip — a single piece of silicon that combines a 20-core Arm CPU, a Blackwell GPU with 6,144 CUDA cores, and 128 gigabytes of unified memory, connected by NVIDIA’s NVLink chip-to-chip interconnect. The whole package delivers up to one petaflop of AI compute in a laptop form factor.\n\nThe number that matters: RTX Spark can run a 120-billion-parameter language model entirely locally, with a context window of one million tokens, without a single byte leaving your machine.\n\nTo put that in perspective: GPT-3 had 175 billion parameters and required clusters of A100 GPUs to run. The model that stunned the world when it launched in 2020 is now approximately the size of what fits in a consumer laptop chip announced this week. The capability that required a data centre in 2020 is coming to a device you carry in a bag in 2026.\n\nPhones can already [run small LLMs](https://spectrum.ieee.org/small-language-models-ai-pharmaceuticals):\n\nIn 2025, slightly more than a third of all smartphones shipped worldwide were capable of running generative AI, and that figure will reach 45 percent by the end of this year, [according to the technology research firm Counterpoint](https://counterpointresearch.com/en/insights/genai-smartphone-share-to-rise-to-45-percent-of-global-shipments-in-2026). By the end of next year, slightly more than half of all smartphones will be able to run a small AI model.\n\nIf Nvidia can already put a 120B-parameter in a laptop, it will only be a few years until phones can run a GPT-3-class model, good enough for almost all consumer needs. Apple and Google own that channel. Owning the channel is better than owning the technology. They will dominate consumer AI, and the other players will garner no revenue from this market.\n\n-\nThird, there is an\n*enterprise* market; all that is left to generate the revenue to service the debts fuelling the AI bubble, lets say [$2T by 2030](https://www.wheresyoured.at/big-tech-2tr/). There are a number of problems that make this unlikely.\n\nFirst, there are very few documented cases of LLM deployment that resulted in enough productivity improvement to cover its *unsubsidized* costs.\n\nSecond, the Trump administration just demonstrated that deploying mission critical systems on AI platforms such as OpenAI or Anthropic means your company can be disabled at [90 minutes notice](https://www.theatlantic.com/technology/2026/06/trump-anthropic-export-control-ai-race/687555/) with no recourse.\n\nThird, this means that companies will have to run mission-critical LLMs on open-weight models on in-house hardware if they are not to be vulnerable to the whims of the US president.\n\nFourth, systems such as RTX Spark show that good enough in-house hardware is likely to become relatively cheap compared to the unsubsidized cost of the AI platforms. In-house systems need much less over-provisioning for demand spikes, and because they aren't shared they don't need to be as fast.\n\nFifth, companies need to balance the productivity benefits (if any) of mission-critical LLMs against the productivity costs they bring. These include a vastly greater attack surface, technical debt from reduced developer understanding of the software, and so on.\n\nThus it seems likely that the hyperscalers and AI plaforms will generate far less revenue than they expect, because they will be restricted to non-mission-critical applications with lower productivity gains, and thus lower pricing power. They will thus be unable to cover the debts they are incurring to build massive data centers predicated on centralized systems dominating (an inflated estimate of) the entire enterprise market.\n\nDavid Wallace-Wells' *Did We Make the Wrong Bet on Big A.I.?* discusses a \"[televised rant](https://www.youtube.com/watch?v=0A3sGymV6kY)\" from Palatir's CEO Alex Karp that supports this argument:\n\nKarp had been softly floating his critique for some time, but the CNBC event looked like a proper coming out. Just one day earlier Palantir had [published](https://x.com/PalantirTech/status/2072114267776491695) a kind of manifesto devoted to what it described as the all-important principle of “A.I. sovereignty.” The central argument: Companies should seek to build their own A.I. tools, not just customize those on offer from the frontier labs. This might mean relying on open-source L.L.M.s rather than the proprietary ones on which the A.I. boom has mostly been built in America, but it would amount to a liberating declaration of independence from Big A.I., which in Karp’s estimation was sucking up much more value than it was generating.\n\n-\n*The Economist*'s *How to turn compute into a financial asset* reports another bad sign (my emphasis):\n\nLocking in a price with a multi-year neocloud contract insures a buyer against compute getting more expensive **but not against it getting much cheaper**. So as businesses around the world spend ever more on compute, they want to be able to hedge against price volatility just as they insure against changes in energy tariffs, interest rates or foreign-exchange movements—ideally in deep and liquid derivatives markets.\n\nTwo startups want to help companies do this, by turning nascent indices tracking compute costs into a futures market. Silicon Data, founded in 2024 and backed by DRW, a trading firm, has paired up with CME Group, which operates large derivatives exchanges. Ornn, created by recent graduates of the Massachusetts Institute of Technology and run from a flat rather than an office just a few months ago, has paired up with Intercontinental Exchange, the parent company of the New York Stock Exchange, to do the same. Both aim to launch compute futures later this year, to be traded on their partner exchanges.\n\n- One characteristic of bubbles is overbuilding the infrastructure. Signs of overbuilding of data centers include that both\n[SpaceX](https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/) and [Meta](https://techcrunch.com/2026/06/05/google-will-pay-spacex-920m-per-month-for-compute/) are now in the business of *rentling their GPUs to the competition*.\n-\nVictor Tangermann's\n*Zuckerberg Admits That AI Is Not Working Out the Way He Imagined* is on-trend:\n\nAs [morale is hitting rock-bottom](https://finance.yahoo.com/technology/ai/articles/mark-zuckerberg-orders-employees-start-123539264.html), his company is [heavily relying on its competitors' AI models](https://finance.yahoo.com/technology/ai/articles/mark-zuckerberg-just-got-rather-134736026.html) to build out its own in-house tools. And despite the many billions of dollars the company has spent in its flailing efforts to keep up in the AI race, even Zuckerberg himself is now acknowledging that progress is nowhere near where he wanted it to be.\n\nAs *Reuters* reports, Zuckerberg admitted during a town hall last week that AI agents in particular aren't progressing as fast as he anticipated, a devastating revelation following enormous layoffs that wiped out [thousands of roles](https://www.nytimes.com/2026/05/19/technology/meta-layoffs-ai.html) at the company.\n\nThe \"trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected,\" he said according to a recording obtained by *Reuters*.", "url": "https://wpnews.pro/news/portents-of-doom", "canonical_source": "https://blog.dshr.org/2026/07/portents-of-doom.html", "published_at": "2026-07-14 15:00:00+00:00", "updated_at": "2026-07-14 18:23:39.944709+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-startups", "ai-policy", "ai-ethics"], "entities": ["Elon Musk", "SpaceX", "Sam Altman", "Dario Amodei", "Torsten Slok", "Panmure Liberum", "Alphabet", "Amazon"], "alternates": {"html": "https://wpnews.pro/news/portents-of-doom", "markdown": "https://wpnews.pro/news/portents-of-doom.md", "text": "https://wpnews.pro/news/portents-of-doom.txt", "jsonld": "https://wpnews.pro/news/portents-of-doom.jsonld"}}