Consensus Capital: National AI Stacks Taking Shape
State of the Future: Dispatch from 8th May 2026: Directionally Very Bad
Hey folks, how you living? You alright? Greetings from Helsinki. I like the sauna, I do not like the karaoke. That’s my take. More travel musings as they come in.
Joined the etn boys again this week who wanted to know why memory stocks were stonking. I can’t wait til Siemens or Mistral buy them out like OpenAI and TBPN to “Accelerate the global conversation around AI.” lol.
Anyway, what do I know about memory stocks? I asked my knowledge base, and it reminded me of this absolute beauty:
“...Or AI is more significant than the Internet and the most impactful technology humans will ever create. In that world, a 50% growth rate is a severe underestimate. In that case, we should be looking for optimisation and efficiency innovations. We continue to closely monitor demand as part of our research. Betting that Nvidia, AMD, and TSMC have their numbers wrong is a tough bet. But what if they have?” Nov 20, 2023!
https://stateofthefuture.substack.com/p/e21-can-we-make-enough-ai-chips
Absolutely nailed it. I am a hunda going to actually start trading, I used to think that Mr Market is pricing in all this correctly, and that an index fund will capture basically all the upside. Aside from those sweet sweet Anthropic secondaries I got. But I dunno man CPO is coming fast and for me Coherent, Marvell and MACOM are just sitting there. BUT 100% NOT INVESTMENT ADVICE. DO NOT LISTEN TO SOME BLOKE IN HIS PANTS ON THE INTERNET.
I jest of course, I am just British. I actually know a lot about this.
—
1. SpaceX is an AI lab now
The big story this week is that SpaceX is an AI lab now. I keep forgetting and the news keeps reminding me. This week’s reminder was that SpaceX wants $55bn for a first-phase semiconductor facility called Terafab, scaling to $119bn across later phases. Intel joined as manufacturing partner on 7 April. The line runs on Intel’s 14A node. USA! USA! USA! Reshoring folks, it’s happening. Customers, per Musk on the Tesla call, are Tesla self-driving, Optimus humanoids, and SpaceX’s own AI data centres.
This is as clean a vertical integration story as you are gonna get. Rockets to launch satellites that beam internet to data centres that train models on chips SpaceX fabs itself, on a process node co-developed with Intel. It’s the Musk Industrial Complex.
Issue 11 closed item 1 with “can’t wait until TSMC’s biggest customer is SpaceX” (I meant biggest customer). Two weeks later SpaceX filed to bypass TSMC entirely. Faster than I thought.
The vertical-integration playbook isn’t new. Google has been training Gemini on its own TPUs in its own datacentres for years. Microsoft has Maia silicon, Cobalt CPUs, Azure, and now MAI models alongside the OpenAI relationship. Apple owns silicon-to-OS-to-services and sort of shipped on-device inference via “Apple Intelligence”. Omg, honestly do you remember this? Jesus christ, it’s so cringe even writing it. What did Apple think that they could co–opt the word “AI”. It’s absolutely mental and not enough people are taking the piss out of Apple for it in my humble opinion.
The hyperscaler-as-vertical-bundle is becoming the dominant shape of AI now. What SpaceX adds is the layer below the datacentre. Rockets, satellites, Starlink ground stations, the chip via Terafab, the model via xAI. Google, Microsoft, and Apple stop at the datacentre door. SpaceX owns everything from the launch vehicle to the inference call. Here’ another call I’ll make: Intel will be acquired by one of these guys in <12 months.
Source: TechCrunch. Background: LFG (for semiconductors) — Dec 2025, where I argued the foundry layer was where the next decade of value capture lived. SpaceX took that further than I had in mind.
2. Big Fund is leading DeepSeek
Meanwhile in China. China Integrated Circuit Industry Investment Fund (the Big Fund, the state chip fund) is “in talks” to lead DeepSeek’s first-ever external round at roughly $45bn. Tencent and Alibaba have FOMOed in. Valuation has roughly doubled in a few weeks; the figure was $20bn in late-April apparently.
And here’s what everybody is missing (can you legit write that even in irony anymore?), DeepSeek’s models are optimised end-to-end for Huawei silicon. I didn’t realise this
So the Big Fund (chips) leading the round (model lab) bundled with the cloud distributors (Tencent, Alibaba) is an integrated industrial play.
It’s happening guys, we have the emergence of a US and China AI stack from silicon > model > hyperscaler. This is the China variant of the bundle I described in A Specific Theory of Sovereign AI back in October. The thesis was that sovereignty isn’t a model layer or a chip layer, it’s the state-level integration of compute, models, and distribution into one industrial “object”. DeepSeek + Big Fund + Huawei + Tencent + Alibaba is that, in cap-table form.
We need to fix up and look sharp here in UK/EU. Consensus Capital I’ve called it in the past. See below:
Source: Bloomberg | TechCrunch. Background: A Specific Theory of Sovereign AI — Oct 2025, the framework piece. Sovereignty as state-level vertical integration, not regulation.
3. Europe wrote a letter
I know. A letter. Ha, ha, ha, Europe, right. Well, maybe not. This week, the CEOs of seven of the continent’s biggest industrial firms: ASML, Airbus, Mistral, Ericsson, Nokia, SAP, and Siemens, published a joint op-ed. They want 3 things: Simpler AI Act. M&A rules that let European companies actually grow, and protection from “subsidised rivals with very strong market penetration in the EU.” Good stuff, we live in a new world, let’s start acting like it shall we.
The letter was published before a meeting with von der Leyen and lands ahead of the Tech Sovereignty Package on 27 May, which will reopen parts of the AI Act and pair it with the Cloud and AI Development Act and Chips Act II. The first formal EU-level definition of “digital sovereignty” is due in the same package. If us Europeans (correct) do anything well, it’s wordsmith. So you just wait for this package. It will be sooo well written, you aren’t going to believe it.
“But, but isn’t this just industry begging for less regulation” Sort of. But read the letter. They’re asking for the Brussels version of what SpaceX and the Big Fund are doing this week. Concentrate capital, shield from imports, and simplify the compliance. They’re asking for permission to build the EU bundle.
Last week’s item 2 was Cohere absorbing Aleph Alpha at $20bn, which I called the sub-scale-incumbent-bailed-out-by-government version of European sovereignty. This letter is a bit different but comes from the same rationale.
Tech Sovereignty Package on 27 May is the test. If it is just some sort of procurement directive plus disclosures, then I am afraid it’s probably over. I think maybe it’s already too late if the scaling hypothesis materialises, but we gotta at least try. I’m thinking just off the dome, no AI:
Silicon: ARM, Infineon, STMicro, Graphcore?
Tools: ASML, Carl Zeiss, Trumpf, Siemens EDA, ASM, IMEC
Networks: Ericsson, Nokia,
Vertical Models: Mistral, Wayve, Helsing?
No leading frontier general model, leading node fab or memory player (FMC soon?). Obviously ton more gaps, but I’m trying to simplify here.
Source: Reuters via Yahoo. Background: How to Invest in AI Sovereignty — Sovereign Albion w/ Andrew Bennett — Feb 2026, the playbook for European sovereignty assembly Lawrence and Andrew worked through. Sharper now that ASML is signing op-eds.
4. MCP at 97 million
Not too many moons ago, Anthropic shipped a tools-protocol called MCP (Model Context Protocol) as a side-project. Just a short 18 months later it has 97 million installs as of March 2026 (so probably 100m+ now) and every major AI provider is shipping MCP-compatible tooling, inc. OpenAI. It’s the default mechanism by which agents connect to external tools, APIs, and data sources. Same broader rev cycle, Anthropic ARR overtook OpenAI for the first time at $30bn run rate vs OpenAI’s $24bn, primarily on enterprise agentic workloads.
Anthropic gave MCP away on purpose. It’s the side-project version of the first Stripe API. Or Android (AOSP for those who know). By making it open and plug-in-anywhere, Anthropic ensured every tool integration in the world routes into Claude as easily as it routes into anything else. Every CRM connector, every database adapter, every SaaS hook written to MCP works on Claude by default and works on the others as a port. Open source is a great strategic tool when wielded carefully.
I remember Gary M tweeting last year that Anthropic was “losing the platform war” to OpenAI’s GPTs. .@Gary, mate.
A few thoughts on where open source could go next. Or another way of saying I hope you aren’t running this startup or backed one that is a nearest neighbour:
Agent identity and authentication. MCP solved how agents talk to tools. It didn’t solve how tools know which agent is calling, which human is behind it, with what scopes, and what audit trail. Today it’s bodged OAuth and service accounts. It’s a shitty UX problem right now. Solve asap pls.
Skills format. Anthropic already runs Skills internally (the SKILL.md system). Open-sourcing it as a portable manifest would do for agent capabilities what npm did for JS libraries. Registry follows. Every Skill written for Claude becomes portable, but Claude is the reference implementation. The most Stripe-API-shaped.
Agent-to-agent handoff protocol. MCP stops helping when agents call each other. Orchestration is a mess of proprietary frameworks (LangGraph, CrewAI, and Swarm). A standard envelope for “agent A delegates task T to agent B with context C and budget B,” with provenance, partial-result streaming, and cancellation, is very MCP-shaped. Google is gesturing at this with A2A.
Source: Anthropic on MCP | Background: You Like AI Agents, You Are Gonna... — Feb 2025, before MCP was a thing. The piece argued the agent runtime would converge on a single protocol layer and that whoever owned it would own the value. Aged well.
I would appreciate any feedback on the newsletter.
If you could indicate directionally good or bad?
—
If you missed it: What Are We Even Counting — Issue 12, last week’s piece on Anthropic at $900bn and the absorbed pure-play AI labs. Sets up most of this issue