Moonshot AI dropped Kimi K3 on July 16, and within about 48 hours an OpenAI executive was publicly warning that models like it could usher in full AI communism. That reaction hould tell you more than the benchmark scores do. K3 is a 2.8-trillion-parameter open-weight model, the largest anyone has released, and on Artificial Analysis’s Intelligence Index it landed at 57, good for a top-four finish and, notably, a hair ahead of Claude Opus 4.8 at 56. It beats Opus on agentic tasks too (an Elo of 1,668 versus 1,600). The full weights are supposed to be public by July 27, free for anyone to download and run. Sitting behind that model is a company whose previous flagship, K2 Thinking, was reported to cost about $4.6M to train, on Nvidia’s export-restricted H800 chips. Moonshot’s CEO, Yang Zhilin, has said that $4.6M number “is not an official” figure and that real training cost is hard to pin down because so much of it is research and failed experiments. But even the skeptical outside estimates land around $5–6M of base compute. Round it however you like, it’s three orders of magnitude below what a US frontier training run is assumed to cost.Now hold that next to the capex numbers touted by the mega caps. Amazon, Alphabet, Meta, Microsoft, and Oracle are collectively guiding to somewhere around $700–725B of capital spending in 2026, up roughly 77% from about $410B in 2025. The Street already expects the group to clear $1 trillion in 2027. Goldman has the four biggest hyperscalers spending something like $5.3T cumulatively between now and 2030. So the obvious question writes itself. If a Chinese lab under a chip embargo can build a model that trades punches with Opus for the price of a nice house in Palo Alto, what exactly is the $700B for? That’s the lazy version of the question, and it’s the one the market keeps reflexively asking every time a good Chinese model ships. I think it’s the wrong question, or at least an incomplete one. The more interesting fact about K3 isn’t that it’s cheap. It’s that Moonshot raised the price. K3 costs $3 per million input tokens and $15 per million output, roughly triple what K2.6 charged three months ago ($0.95 and $4.00), and right on top of where Anthropic prices Claude Sonnet. On a per-task basis K3 runs about $0.94, versus roughly $1.04 for GPT-5.6 Sol and $1.80 for Opus 4.8. For two years the story about Chinese models was that they’d collapse the price of intelligence to near zero and take everyone’s margin with them. And at the very bottom of the market, that’s still happening as DeepSeek V4 Pro will serve you a task for about four cents. But the best Chinese open model just chose to price like a Western incumbent instead of racing to the floor. That’s a signal about the actual economics and worth looking at.The reason is that the cheap part of a model is building it and the expensive part is serving it at scale, reliably, to millions of concurrent users. A 2.8T-parameter mixture-of-experts model is not cheap to run even if it was cheap to train, and (this is the part that matters) China can’t get enough of the memory it would need to serve one to the world. K3’s price hike is Moonshot telling you that frontier-adjacent inference has a real cost floor, embargo or no embargo. There’s a quality tell too, K3’s benchmark gains came with its hallucination rate climbing from 39% to 51%. It got more capable and less trustworthy in the same release. The true frontier, Fable 5 and the top GPT-5.6 Sol configurations, is still measurably ahead, and it’s ahead precisely on the reliability dimensions that matter most for the highest-value work. Open weights are eating the good-enough middle of the market, they are not yet eating the frontier, and the frontier is where the pricing power lives.Don’t mistake any of this for the open-source challenge being soft. By OpenRouter’s own accounting of 100 trillion routed tokens, Chinese open-weight models went from roughly 2% of open-weight usage to about 61% in the span of eighteen months. Alibaba’s Qwen passed a billion cumulative downloads on Hugging Face and overtook Meta’s Llama as the most-downloaded open model family on earth. Something like 40% of new open LLM derivatives are now built on Qwen. The Chinese open cohort is deep, not a one-model story. DeepSeek V4 Pro (1.6T parameters, MIT-licensed, the cost leader). Zhipu’s GLM-5.2 (744B, MIT, strong on long-horizon coding). MiniMax’s latest, priced at $0.30/$1.20. Qwen underneath all of it as the default base everyone fine-tunes. Kimi K3 now sits at the premium end of that group. What’s striking is the license: DeepSeek and GLM are both MIT, which is about as permissive as it gets. This is a genuine, well-supplied, permissively-licensed frontier-adjacent stack that any startup, any enterprise, any government can pick up for free.The quiet loser here is the American open-weight strategy. The most-downloaded open models in the world are Chinese now, and Llama is the one that got lapped. If you believed open weights were going to be a US-led soft-power play, that thesis is not aging well. But China’s models are good and US compute demand falls are different claims Here’s the leap the bears keep making, and it’s the load-bearing error. They go from efficient open models exist straight to so the buildout is a waste. Those are not the same statement, and the gap between them is where the whole investment case actually lives.Start with what happened the last time. In January 2025, DeepSeek R1 was supposed to be the event that ended the capex supercycle. Nvidia fell about 17% in a day. The logic was identical to today’s: China proved you don’t need the compute. What actually happened over the following eighteen months is that inference demand exploded. OpenRouter’s token volume went from around 5 trillion tokens a week to over 20 trillion or roughly 4x. Total inference grew faster than per-token efficiency improved. Nvidia is now worth something like $5.1T, up roughly 50% year over year. The DeepSeek moment was, in hindsight, a buying opportunity, because cheaper intelligence didn’t shrink the market it expanded it. That’s Jevons paradox, and it’s not a cute analogy here, it’s the actual observed pattern in the token data. So the first-order efficiency kills demand case has now been run as a live experiment once, and it lost. I’d want a strong reason to bet on it a second time, and I don’t think K3 is one. If anything, a great free open-weight model increases total inference as more people build more things and run more tokens through them, and most of those tokens get served on Nvidia silicon somewhere.Then there’s the fact almost nobody prices in: China physically cannot serve these models to the world. The binding constraint on Chinese AI hardware is no longer SMIC’s ability to etch logic but is high-bandwidth memory. Domestic HBM from CXMT is projected at roughly 2.0 million stacks in 2026, which is enough to finish only about 250,000–300,000 Huawei 910C-class accelerators for the entire year. That is a rounding error against what a single US hyperscaler installs. China can design and train a frontier model on a shoestring. It cannot deploy frontier-scale inference capacity at global volume, because it can’t get the memory. Which means a world full of excellent Chinese open weights is still, in large part, a world where those weights run on Western hardware in Western and allied data centers. That decoupling is the single most important thing to internalize. Kimi K3 is a model-layer disruptor and is not an inference-capacity disruptor. It pressures what you can charge for a given tier of intelligence, it does very little to reduce how much compute the world consumes and the compute consumption is what the $700B is actually buying.The Real RiskNone of this means the capex is safe, it means the danger is in a different place than China proved it’s unnecessary. The real risk is boring and financial, return on capital and the depreciation clock. And here the numbers have genuinely turned less comfortable. Look at what the spending is already doing to cash flow, using reported figures. Amazon’s free cash flow fell from about $33B in 2024 to under $8B in 2025 and that’s before the $200B 2026 capex guide lands. MS now models Amazon at roughly negative $17B of free cash flow this year. Oracle is the sharp end of the whole trade, on about $56B of capex in its latest fiscal year it posted roughly negative $29B of free cash flow, swinging from negative $5B the year before. Meta’s net income actually slipped year over year, from $62.4B to $60.5B, as depreciation and expense growth caught up with it. Across the group, capex is now running near 100% of operating cash flow, versus a historical average closer to 40%. Epoch AI projects that aggregate cash capex overtakes operating cash flow sometime in Q3 2026. That’s the crossover the bears have been waiting for, and it’s basically here. Then layer on Burry’s argument, which is the most intellectually serious version of the bear case. He’s not saying the chips are useless. He’s saying the accounting is a lie of omission: hyperscalers depreciate GPUs over five to six years while the real economic life is closer to two or three, which means reported profits are overstated by understating depreciation, by his math, around $176B across 2026–2028. He calls it “token maxing” and, borrowing Galbraith, the “bezzle.” If he’s directionally right, a lot of the earnings that justify these multiples are borrowed from a future writedown.The bull rebuttal is that the revenue is finally showing up. By Exponential View’s tally, AI infrastructure now generates about $1.19 of revenue for every $1 of depreciation, having crossed above 1.0 for the first time in the past year. Sequoia’s David Cahn has framed the same tension as a running revenue gap between what’s been spent and what’s been earned back. The honest read is that the industry is in a race between a depreciation schedule that’s arguably too generous (although recent evidence shown that is not the case) and a revenue ramp that’s genuinely steep. Nobody actually knows which curve wins, including the people spending the money. That’s the real debate and not did Kimi make the GPUs pointless. It’s will the cash come back before the chips are written down and need to be replaced. Where the pain landsIf the model layer is where open weights bite and the compute layer is where demand keeps compounding, then the exposure map is not uniform. Nvidia is the least exposed to the Kimi story, and the clearest proof is in its own numbers. It just wrote off effectively its entire China business and the H20 is dead after Beijing told domestic firms to buy local, the H200 trickles in under a 25% US export tax, and next-quarter guidance assumes zero China data-center revenue, erasing what analysts once pegged near $20B a year, and it still grew revenue 85% to $81.6B last quarter, data center up 92%, gross margin at 75%, with the next quarter guided to $91B. When you can delete the world’s second-largest market and still not build chips fast enough, the binding constraint is demand, and Jevons plus China’s HBM ceiling means more good open models point more inference at your silicon, not less. Consensus has Nvidia printing north of $200B of free cash flow this fiscal year on the way to ~$280B, at ~22x forward and ~16x on FY2. You are not paying a bubble multiple for the one company whose demand actually benefits from open-weight proliferation. The real risk to Nvidia isn’t Chinese models; it’s the plumbing underneath its own demand. Customer concentration is climbing fast as three customers were 21%, 17%, and 16% of revenue last quarter, and the top hyperscalers are roughly half of data-center sales. And some of that demand is vendor-financed: Nvidia now holds around $50B of equity spread across its own customers, including a $30B stake in OpenAI (quietly cut down from the headline $100B “letter of intent” when the deal was restructured in February) plus Anthropic, xAI, CoreWeave, and Intel. Critics from Mizuho to Bernstein to Michael Burry (who is short and calls Nvidia the Cisco at the center of the bubble) argue this is pre-funding the purchase of your own chips and Nvidia’s rebuttal was a memo amounting to we are not Enron. Both things can be true at once. Demand is genuinely parabolic, and it is concentrated in a handful of customers Nvidia is helping to fund. Google looks like the best-positioned hyperscaler, and the past few months strengthened the case. It runs its own TPUs, so it isn’t hostage to Nvidia’s pricing, and it has now turned that edge into a merchant business: the Ironwood generation (TPU v7) is going outside Google at scale for the first time, anchored by a deal to put up to a million TPUs and more than a gigawatt of capacity into Anthropic’s hands, worth tens of billions and routed partly through Broadcom. Google Cloud grew 63% last quarter to $20B at roughly a third operating margin, and backlog nearly doubled to $462B. Gemini 3.1 briefly topped the intelligence leaderboard when it shipped in February. Add the cheapest blended cost of compute in the group and you have the one hyperscaler monetizing the buildout two ways, its own models, and other people’s workloads on its own silicon. Two honest caveats keep it from being a free lunch. The frontier moved past Gemini again: by July it sits just behind the Fable 5 / GPT-5.6 / K3 cluster rather than on top, and Google’s flywheel depends on staying near the front. And that eye-popping 63% cloud growth and $462B backlog are partly mechanical, booking large TPU hardware sales inflates both, so it isn’t all pure services demand. Even Google feels the capex: consensus sees its free cash flow compressing from $63B in 2025 toward ~$20B in 2026. But that’s a dent, not the swing to negative Amazon and Oracle face, and it stays firmly cash-generative. At ~24x forward it’s the name where the spend most plausibly earns its keep — and now it sells the picks it used to hoard.Meta just got a lot more interesting, and it’s the cleanest live illustration of everything above. The setup is 2026 capex guided as high as $145B, up 73% from $72B, against a business that threw off $41B of free cash flow last year (down from $49B) but whose net income actually slipped year over year as depreciation caught up and whose free cash flow consensus now expects to hit roughly breakeven in 2026 and turn slightly negative in 2027 once the capex fully lands. Zuckerberg is spending Microsoft-scale capex to chase superintelligence, and for most of the past year the honest investor question was blunt: what does a social-media company do with that much compute? In July, Meta gave a revealing answer and per Bloomberg, it’s building Meta Compute, a business to rent out excess GPU capacity, either as raw silicon (competing directly with neoclouds like CoreWeave) or as hosted access to its own closed model, Muse Spark, in an AWS Bedrock style. The stock jumped 8.8% on the July 1 report and added north of $100B in market value. CoreWeave and Nebius fell 10.8% and 12.4% the same day. That one event is a two-sided tell, and both sides matter. The bullish side: Meta apparently has so much demand for compute that it can profitably sell the overflow, which validates the Jevons thesis and hands the capex a second revenue leg with nothing to do with ads. The side management won’t say out loud: if you’re standing up a business to sell excess compute, you bought more than your own products can absorb. Meta is quietly converting part of a superintelligence moonshot into we’ll rent it out like infrastructure if the moonshot takes a while. It’s a hedge against its own capex bet, packaged as an expansion, and the market liked the hedge enough to pay $100B for it. It also makes my neocloud caution worse, not better. A hyperscaler with its own cash flow, its own power contracts, and a near-zero cost of capital stepping into merchant compute is an existential problem for pure-play neoclouds that borrow to buy the same chips. Nebius dropping 12% wasn’t a tantrum, it was the market correctly pricing a new, better-capitalized competitor. If Oracle is the levered top of the buildout, Meta Compute is the thing that can crush the layer beneath it. The forward question I’d actually track is which version of Meta Compute this becomes. Selling raw GPU is a low-margin, bare-metal commodity business, useful for absorbing capex, not for creating much value. The high-value version is hosted Muse Spark, and that only works if Meta’s closed model is good enough that developers pick it over OpenAI, Anthropic, and Gemini, for which there’s no evidence yet. Note the strategy reversal buried in here: the company that tried to make open weights a US soft-power play with Llama, then watched Qwen lap it, is now pivoting to a closed, monetizable model plus a cloud to sell it. That’s a quiet admission about where the money actually is. At ~20x forward, after the Meta Compute pop, the stock is no longer obviously cheap, so you’re partly paying for optionality that is, for now, mostly bare-metal. I’d still own it over Amazon and Oracle on cash-flow quality, but I’d want proof that Muse Spark can hold paying developers before crediting Meta Compute as a real second business rather than an elegant way to keep the capex line from looking reckless.Amazon has the demand and the silicon, but the cash-flow optics are about to get ugly. AWS reaccelerated to 28% growth last quarter, its fastest in fifteen quarters, at a 37.7% segment margin, and its custom-chip story is the most credible after Google’s: Project Rainier now runs more than a million Trainium2 chips training Anthropic’s models, Trainium3 is shipping at roughly 4.4x the prior generation, and Amazon is targeting about a gigawatt of its own silicon online by year-end. It has put $13B into Anthropic with up to $20B more committed, and Anthropic has pledged over $100B back to AWS. The catch is the price tag: $200B of 2026 capex, well above the ~$147B the Street had modeled, which is why the stock dropped 8% on the guide. Consensus now takes Amazon’s free cash flow to roughly negative $13B in 2026 before recovering to positive in 2027, and at 28x forward it carries the richest multiple in the group precisely as its cash flow goes underwater. And here’s how circular the whole complex has become: Amazon also committed $50B to OpenAI’s latest round, bankrolling the primary rival to the Anthropic bet it is simultaneously bankrolling. Own it below Google on multiple and cash flow, above Oracle on quality.Oracle is the purest and most levered expression of the entire trade, and the name where the depreciation-and-ROI air pocket is most likely to actually detonate. On paper the backlog is staggering (remaining performance obligations exploded 363% to $638B), but roughly half of it is a single customer, OpenAI, via a reported $300B five-year contract, and OpenAI is itself losing something like $14B a year. Only about 12% of that backlog converts to revenue in the next twelve months. To build for it, Oracle spent $55.7B of capex in the year just ended (up from $21B), guided next year’s gross capex to $90–95B, and is running deeply negative free cash flow, very roughly negative $24–29B last year, with consensus and S&P modeling something like negative $42–46B this year and negative $41B the year after. The GPU cloud it is building carries gross margins of only ~14–16%, per leaked internals, and it is reportedly losing money outright on some rentals. Funding the gap has pushed total debt to about $167B, and on July 9 S&P cut Oracle to BBB-, one notch above junk, explicitly citing the OpenAI concentration; its credit-default swaps blew out to the widest since 2008 before easing. The stock just had its worst week since the 2001 dot-com bust and sits nearly 60% off its high. At ~16x forward it looks cheap only if you assume the backlog converts and the depreciation math holds, the two things most in doubt. This isn’t a Kimi story in the slightest. It’s a balance-sheet story, and it’s the first domino if sentiment on capex returns cracks.The closed frontier labs, which public investors mostly reach through Microsoft/OpenAI, Google/Gemini, and Amazon-and-Google/Anthropic, are where Kimi does its real damage, and it is a margin story, not a demand one. A free, MIT-adjacent model at Sonnet parity caps what anyone can charge for good-enough intelligence, and the price war is already live: OpenAI, xAI, and Meta all cut token prices this year with the Chinese open models setting the floor. That is what Dean Ball of OpenAI was really reacting to when he called an open-weight-dominant world “full AI communism” and a “dystopian hellscape,” the sound of an incumbent watching the bottom drop out of the commodity tier of his own price list, even as he conceded Kimi is “a very good model.”But strip the theatrics and Ball’s post reads less like analysis than like a lobbying document. Start with the word itself, Communism, is a strange label for software that a private company built with private capital and then gave away for anyone to run however they like. Open weights are the most decentralized, permissionless layer in the entire stack, the opposite of state control. What he is actually describing when he warns that intelligence becomes a cheap public good isn’t communism, it’s competition, doing exactly what competition is supposed to do, dragging the price of a commodity toward its cost. He just doesn’t like where that leaves his employer’s margins. The decelerationist line is the real tell. The complaint is that open models are dangerous because they “reduce the incentive to invest in frontier AI.” Restore the self-interest and it reads plainly, cheap competitors compress our returns, so they must be bad for progress, so someone ought to do something about them. That is the oldest move in the incumbent playbook. You reframe a threat to your business as a threat to the country, wrap it in a China narrative, open weights are Chinese, Chinese is adversary, adversary is security risk, and export-control-style curbs on open models start to look patriotic instead of protectionist. It’s regulatory capture with a flag on it. The irony is that Ball made his name as a light-touch, permissionless-innovation writer before he joined OpenAI; the principle lasted right up until the most permissionless innovation of all began undercutting the house. In the same post he calls Kimi “very token hungry” and grants that it’s “not obvious it’s actually that cheap to run.” If it is expensive to serve, it isn’t collapsing anyone’s economics and there is nothing to protect against. You can argue open models are a pricing threat, or that they’re costly to run, but you can’t really argue both, unless the logic was never the point, and the audience was always Washington.For what it’s worth, the frontier premium is holding with no protection at all. Anthropic just posted its first profitable quarter at north of 80% API gross margins, straight through the price war, which rather undercuts the notion that anyone’s incentive to keep building has been destroyed. OpenAI is the more unsettling case, but not because of Kimi, it’s the economics, around $25B of revenue against a reported $14B annual loss and a confidential IPO filing said to target north of $1 trillion. The read across the labs is clean enough for now, whoever owns the genuine frontier and a distribution channel keeps the margin and the me-too middle is exactly where an open Chinese/US model takes the business.The other side of the API: cheaper intelligence and the software wreckageEverything above is about the people building and selling the compute. But collapsing inference cost has a second-order effect one layer up, and it points the opposite way. If intelligence is getting cheap and abundant, that is a cost windfall for the companies that consume it, and the most beaten-down corner of the market right now is exactly the group that should benefit, application software.The de-rating there has been violent. In Q1 2026 the software complex had what people started calling the SaaSpocalypse, the IGV software ETF fell about 24%, its worst quarter since 2008, triggered when Anthropic’s agentic Claude Cowork crystallized the fear that autonomous agents simply replace per-seat software. The wreckage is specific and severe. Adobe is down roughly 43% from its peak to under 10x forward earnings, a multiple it hasn’t seen in years, on fear that generative tools commoditize Creative Cloud, even as its AI-first ARR runs past $500M and triples year over year. Salesforce trades around 12x forward against a software-industry average closer to 26x. Workday sits near 13x, HubSpot got cut roughly in half. By a16z’s measure, software multiples have compressed to 2014 levels on forward free cash flow, the lowest premium in over a decade, and the drawdown showed almost no correlation with revenue growth. This was a fear trade based on terminal risk, not a fundamentals trade based on current numbers.The bull case is the same Jevons logic that runs through this whole piece, just pointed at a different beneficiary. Inference cost is falling something like 10x a year, down roughly 95% in two years for a fixed level of capability, while total tokens consumed keep exploding. That is a straight gross-margin tailwind for any software company embedding AI into its product, and it undercuts the premise of the sell-off. The sharper version, which a16z laid out in “Good News: AI Will Eat Application Software,” is that the value in software was never the code, it was the proprietary data, the system-of-record status, the distribution, the workflow lock-in, the “process power.” Commoditized cheap models don’t destroy those moats, they let the incumbents that own them bolt on AI features at almost no marginal cost. And the monetization is no longer hypothetical: Salesforce’s Agentforce is already at a $1.2B run-rate, up more than 200% year over year, and Intuit is rolling out usage-based AI pricing this August on top of a TurboTax Live business growing 36%. The names the market marked for death are, so far, selling AI successfully.The honest tension is that this tailwind is selective, not a sector beta. Cheaper unit inference doesn’t automatically lift margins, because Jevons pushes the absolute AI bill up even as the per-token price falls, and AI-native gross margins (50–60%) run well below mature SaaS (70–90%). The bear mechanism is real too: if ten agents do the work of a hundred seats, seat-based revenue compresses even when the software is still used, and the most aggressive version of the argument, that agents eventually disintermediate the application and ERP layer outright, can’t be waved away. The market has already sorted the sector into two buckets on exactly this logic. The AI beneficiaries, data and observability names like Snowflake (~139x forward), Datadog (~107x, back at all-time highs), and MongoDB (~51x), never stayed down; the market happily pays up for the consumption-upside framing. The AI victims (Adobe, Salesforce, Workday) are where the cheap multiples live. The contrarian trade isn’t buy software. It’s buying the victims that actually own the data, the system of record, and the distribution, where cheaper intelligence is a mispriced tailwind, while avoiding the thin AI wrappers and pure seat-based licensors that genuinely are exposed to headcount deflation.Notice that this is the mirror image of the capex trade. If the capex bears turn out to be even partly right that the buildout’s returns disappoint, the reason will be that intelligence got cheap and abundant faster than it could be monetized at the frontier, which is precisely the windfall for the companies sitting on the consuming side of the API. The same cheapening that pressures the frontier labs’ pricing power is what should re-rate the best of the beaten-down software names.