The Once and Future Kimi

Lawrence Lundy’s State of the Future: Dispatch from 17 July 2026

“The perilous season is middle age, when a false wisdom tempts them to doubt the divine origin of the dreams of their youth.” Elizabeth Palmer Peabody

“If this time really is different, your biggest mistakes will be applying what you’ve already learnt” Lawrence Lundy-Bryan

“Boring isn’t it? Just staying in and watching Peak Practice with your life” Big Keith

---

Two newsletters on the trot. You are lucky. You might not have even got to the HBM one from yesterday. I’ll keep this to 1 story so you can read it in one sitting.

---

Kimi

Key change to prior: Chinese open source is frontier-competitive, <6 months behind, and it looks like deliberate CCP strategy.

New Chinese frontier lab release alert. If DeepSeek caused somewhat unnecessary panic, with Kimi, trained on an old Macbook or whatever was claimed, I wouldn’t say “panic” but maybe, “take your enemy more seriously”.

First, facts. Moonshot released Kimi K3 this week.

• Size: 2.8T parameters with only 16 of 896 experts firing per token, 1M context. 896 experts don’t fit on one chip, so they’re scattered across 64.

• The router picks 16 experts for each token, and those 16 live on 16 different chips.

• Performance: It wins some benchmarkts outright, like game dev and UI on the early tests, but trails Fable 5 and GPT-5.6 Sol overall.

• Price: $3 in, $15 out. We’re talking Sonnet money. The Chinese playbook used to be 6 months behind at a tenth of the price.

• No longer “6 months behind”, but also not “10% of the cost” either.

• Open-source not closed.

Software vs hardware

When a model writes, it re-reads everything in the conversation so far, kept basically in a scratchpad (the infamous KV cache (key-value cache) you might have heard about) in HBM. The entire scratchpad gets read for every token, so longer chats = higher memory bills = why HBM is sold out aka The Memory Trade/Koreans at the ATM meme.

The Chinese labs are engineering around this. DeepSeek’s MLA (Multi-head Latent Attention) compresses the scratchpad before storing it. And they use a technnique called sparsity which means not waking the whole model. K3 is 2.8T parameters but only 16 of its 896 experts fire for any one token. And Kimi Delta Attention goes even further: it sort of throws the scratchpad away and keeps a fixed-size running summary, deciding channel by channel what to forget as it goes. Clever stuff. They claim 2.5x better scaling efficiency than K2.

This is the inevitable consequence of being compute-constrainted (same story as litho. SMIC got to 7nm-class logic on DUV because it couldn’t buy EUV). (Biggest bear case against NA-EUV and ASML btw, is SMIC gets to 3-5nm with DUV somehow) We can now say for certain, export controls are accelerating the algorithmic efficiency that erodes the compute moat they were intended to protect. I wouldn’t say unexpected consequences.

I won’t overindex on one new architecture but we should still update on this. Less bandwidth per token, sure. But more memory overall, because sparsity is what lets you build 2.8T parameters in the first place, and all of it has to sit somewhere fast. And much more interconnect: Moonshot recommend running K3 on a supernode of 64+ accelerators, because each token gets routed to 16 experts scattered across those chips, every layer, and they all have to talk over an NVLink-class fabric to keep up.

As the frontier incorporates these algorithmic innovations, it’s all open source, then what you buy memory FOR changes: less for bandwidth, more for capacity. The premium HBM charges for bandwidth gets harder to defend. This may be the starting gun for the move to the Interconnect Trade. Photonics people look alive.

So why does China keep giving frontier models away?

The most interesting question in geopolitics today: why does the CCP continue to give away frontier AI for free?

John Loeber’s thinks: open weights commoditise OpenAI and Anthropic’s core asset, and if the US labs’ primacy looks uncertain their teams splinter into startups built on, well, open weights. Arguably the schism already happened, OpenAI begat Anthropic begat Thinking Machines.

Teortaxes thinks: proliferate capability you can’t inference at scale anyway, and the US burns its compute lead policing models instead of extending them.

i think the simple version is the true one. This is CCP-sanctioned commoditisation. Nothing strategic comes out of China by accident. The best outcome for China is that if they can’t decisively win at the frontier then best option is to get as close as possible and get the rest of the world using it for free instead of the US models.

It also tells you the odds of Kimi getting locked down later are small, you don’t shut down the thing that’s working.

The US never charged for GPS, with the standard comes the power (see Huawei and 6G btw) and China is running the same playbook at the model layer, DeepSeek, Moonshot and Qwen keeping each other honest.

An interesting angle to this is that open source was the cypherpunk commons, Linux, Apache, Python, Kubernetes, RISC-V, all insurance against concentrated power. In 2026 the biggest sponsor of free frontier software is China, running it as strategy. The commons is a weapon now, wielded by the least libertarian actor on earth against the country whose hobbyists invented it. Open is indeed better than closed. But we've flipped sides. Maybe we are now the baddies?

Potential options

So what does the US do? You can reach for the ban option. But weights are torrents under Apache 2.0, and fine-tuning confuses matters. And if you do that you are only banning US companies. So you end up with more expensive American AI, less profitable American companies, and the rest of the world’s developers will just use chinese models. Sub-optimal.

Or you can try and win at the open weights game. America’s entry is sort of gpt-oss, roughly o4-mini class, still the current distribution 11 months on. More a gesture really. But as of this week, we have Thinking Machines with “Inkling”. 975B parameters, 41B active, open weights, but its “not the strongest overall model available today, open or closed.” Same architecture family as K3 at a third the size. But Thinking Machines isn’t aiming at the frontier, its lead gen for Tinker, their fine-tuning platform. Note, they used Kimi K2.5 to generate post-training data. The US open-weights play is partly built on the Chinese open-weights model.

Notice the open source came from the lab with no capex. When you spend upwards of a trilli on token factories, you really have to charge for them. We’re talking real money now. But capex only gets you halfway, because the man with the most capex on earth opens more weights than anyone. So look at who can actually do this. You need 2 things to be America’s open weights answer.

• You have to be willing, meaning no token revenue to cannibalise, and

• You have to be able, meaning a cluster big enough to train a frontier model.

Nobody in America has both. Thinking Machines is willing, so it opened, but not a frontier model. Musk is able, obviously, and he opens more weights than anyone, Grok-1, Grok 2.5, Grok 3 pledged, v8-small promised by year end. But it’s always n-1, and n-1 American loses to n Chinese every single time.

Which leaves OpenAI and Anthropic, able and unwilling. Unless of course the Government forced them to release open weights, but then you wipe multiple trillions off stock market valuations and at this point, with basically the US stock market and economy propped up by datacentre build out, probably a recession. Sub-optimal.

The best hope would be if the CCP shot itself in the foot and closed the next gen of models. But the CCP have proven themselves pretty strategically savvy around this whole technology game.

Kimi really brings it home for the US.

1. Ban Chinese open source, somehow, and you make your AI more expensive and less profitable and cede the rest of the world to China.

2. Don’t, and the world’s intelligence substrate is Chinese made, someone else’s values, someone else’s standards, with no-off switch or leverage.

If geopolitics and economics weren’t in conflict before, there certainly are now.

Wot wud u do if u woz us president rn? answers in comments

---

STATE OF THE FUTURE ALMANAC · 1950-2050 (formally The Book)

• [assumption: frontier-llm-scaling] ↓ 69 to 64. The assumption: gen-6 models in 2028 need $100bn training clusters and 10GW of power, and gen-7 by 2030 runs to $1T of capex, so brute scale stays the price of entry. The gen-5-ships-by-2026 leg already resolved correct. So the 2030 bet: if 2.5x-per-generation architecture gains are real and they diffuse, the mega-cluster maths weakens. I expect this isn’t the end of algorithmic and compression techniques but also Jevon’s Paradox fights the other direction.

• [assumption: model-commoditises-value-moves-up] ↑ 71 to 76. The assumption: raw model capability commoditises, so durable value moves up to the scaffolding and application layer and down to the chips, with the labs climbing into applications themselves. A 7th near-parity supplier, the first Chinese lab confident enough to charge virtually full price, and now a strategic reason to expect cheap frontier intelligence to keep arriving. Strong data.

• [assumption: sovereignty-semi-bifurcation] → 71, unmoved. The assumption: semiconductor supply splits into two ecosystems by 2030, a Western one and a Chinese one, with design IP, equipment and capacity that no longer fully cross, locked in by 2028. Open weights feel like they should dent it but they don’t, because the page is about silicon and tools, which can’t cross the wall, and weights aren’t silicon. What the week adds is Teortaxes’ point that ending Chinese open-sourcing is only purchasable with semiconductor tool relief, which ties the model race straight back to the export-control regime.

• [theme: hbm-free-inference-architectures] → 68, unmoved, and I nearly marked it down. The theme: HBM stays scarce and expensive through about 2028, and that scarcity window, rather than long-run economics, is what makes HBM-light AI silicon worth backing. The bet is on a timing window which is risky business in VC. K3 is less bandwidth per token meaning less need for HBM. Except it needs a 64-accelerator supernode to run, and sparsity is what makes a 2.8T model economic in the first place, so the amount of HBM climbs even as bandwidth per token falls. Those two cancel. So volume holds but what you’re paying for shifts, capacity not bandwidth.

  • NEW THEME: Interconnect-as-next-leg-up. More to come for paid subscribers.

---

What else I’ve been reading

• Musk bought the power company. APR Energy, basically trailer-mounted gas turbines, over 1GW, it cost an implied $1bn. The turbines were already powering Colossus, but Musk is vertically integrating all the way down. Everyone else signs 20-year nuclear purchasing agreements, he bought generators. Electrek

  • [assumption: power-becomes-binding-constraint] ↑ 90 to 92. The assumption: from 2027 the cap on how big AI datacentres get is electricity and heat rather than chips or memory. Buying the turbine company outright is about as literal as confirmation gets. Only 2 points because there was never much room left above 90.

  • [assumption: energy-buildout-keeps-pace] ↓ 18 to 14. The assumption: generation, baseload and grid build out fast enough that energy never hard-caps the datacentre build through 2030. Already a weakest card, and the one i’m closest to calling broken. The richest man alive resorting to trailer-mounted gas because the interconnection queue runs in years is not a system keeping pace.

• Oratomic raised $300m at Series A for a new type of quantum computer. 3 months out of stealth, Khosla’s largest first ticket ever. The claim is fault tolerance at 10-20k neutral-atom qubits rather than millions. Add OQC’s £260m, Quantum Motion’s $160m and IBM re-upping $10bn this week, a billion dollars of private quantum money in 10 weeks. Still really don’t know how to price all this quantum money, feels SPAC-Y with no clear pathway or timeline to fault tolerance (truthfully). But obviously the prize is worth a bet or two every few years, I suppose. TechCrunch

  • [thesis: quantum-computing-modalities] → 52, unmoved. The thesis: no qubit modality has won yet, so the fundable bet is the enabling layer every machine has to buy, the algos, the lasers, the photonics, the cryo-control, rather than the machine itself. Neutral-atom, superconducting and silicon-spin each raised a mega-round inside 10 weeks, which is the premise holding.

  • [theme: quantum-sensing-market] no score on this one, it’s a space i’m watching rather than a bet i’ve placed. The open question: whether the venture-scale outcome is the silicon in the sensor, the readout and control, or the full system, which leans defence procurement and capital heavy. QuantumDiamonds’ €200m for chip inspection is the near-term-revenue-is-sensing read. Maybe comms too.

---

And that’s it, bugger off.