This week's material points from several unrelated venues to the same thing: compute is completing its turn into heavy industry. GTC Taipei talked power. Computex talked materials. SpaceX put a price on the story with an IPO filing. Indigo's call this week: when compute becomes real estate, the grammar of valuation has to change with it.
2026.05.31 — 2026.06.07 · Once a week. Spot the signals. Recalibrate.
This Week's Signals
Several things happened at once within a single week. At GTC Taipei, Jensen Huang pushed the AI factory story a step further, with the unit of measure moving all the way to gigawatt-scale power. On the Computex floor, the bottleneck kept moving down the signal chain, into materials like copper-clad laminate and quartz fiber. SpaceX announced its IPO (initial public offering) terms: $135 per share, with the pricing date set for June 11, 2026. The same week, Anthropic published a report admitting that AI is already accelerating the development of AI itself. On June 4, Canada launched a national AI strategy. Indigo's posts on X this week covered nearly all of it.
Read as five separate news items, they each file into their own drawer. But Indigo's reading puts them on one curve: AI is going heavy-industrial, and the scarce link is shifting from algorithms all the way to power, materials, and land. And once execution is fully automated, and neither agents nor compute is scarce anymore, the truly scarce thing becomes people — insight is embodied experience (experience that lives in the body and in lived history, and cannot be learned from text), and AI has no body, so that layer is out of its reach. His most widely shared post this week (145 likes) speaks exactly to this human boundary: "These past few days I've been advising many friends: choosing a free place to live matters more than anything. 'Freedom of information, freedom of movement, freedom of money, freedom of expression.' Stay constrained long enough and your mind gets institutionalized too" (original post). Once machines take over execution, a mind that has not been institutionalized, plus real embodied experience, are the only two assets an individual has left.
The takeaway in one sentence: when compute becomes real estate and execution becomes a commodity, pricing power flows to only two ends — the side that has locked up physical capacity, and the people who supply the judgment machines cannot sign off on themselves.
Trends
#01 From GPUs to G-Megawatts: Compute Goes Heavy-Industrial
The densest line of signal this week is infrastructure. After watching Jensen Huang's talk at GTC Taipei, Indigo wrote: "We have gone from chasing GPUs to chasing 'G-megawatts.' The NVIDIA DSX platform provides a complete solution for AI factories. This is the 'heavy industry' of the intelligence era" (original post). G-megawatts means gigawatt-scale (billions of watts) power capacity — the unit of competition has switched from chip counts to power capacity. An AI factory is a data-center system that mass-produces intelligence as a finished product. He also compared NeoCloud (a new kind of cloud company that builds AI data centers and rents out compute like property) to the silicon-based real estate developers of a new era. He gave the magnitude at an offline talk in Hong Kong: over the next 4.5 years, the world is expected to add 150GW of compute capacity, roughly $7.5 trillion of investment — close to 5% of global GDP, more than military spending. With money at that scale, compute is no longer software; it is real estate. And real estate is priced in tenants, rents, and capex — not user-growth multiples.
#02 SpaceX's $135: The First Public Quote for the New Story
Indigo took the offering materials apart line by line: SpaceX has built a dedicated website for its IPO… priced at $135/share, pricing date June 11, 2026… 2025 revenue $18.7B (+33% YoY)… GAAP net income -$4.9B, mainly due to heavy AI investment… This is a story that uses launch-cost advantage as the moat, expands sideways into connectivity and AI, then uses the AI narrative to lift the valuation ceiling — but current profits are entirely swallowed by AI capex, and the valuation anchor rests almost completely on distant margins coming true
(original post). He said plainly the price is not cheap. GAAP means US Generally Accepted Accounting Principles; on that basis, the company is still losing money today. The bigger issue is positioning: in his Hong Kong talk, he pointed out that SpaceX is going public precisely as a NeoCloud — its data-center business, Colossus, added $15 billion in revenue in one year, roughly equal to the rocket business's total revenue over twenty-plus years, while Anthropic's $45B contract serves as the anchor tenant (a large customer that locks in long-term demand upfront so a project can be financed on cash flow). This is not an aerospace company going public. It is the first time the market has publicly priced the AI heavy-industry story.
#03 The Mirror for Recursive Self-Improvement: Is the Verifier in the Loop
Anthropic published a report this week, When AI Builds Itself,
admitting that recursive self-improvement (AI participating in and accelerating the development of the next generation of AI) is already happening internally: over 80% of production code is written by Claude, and research speedups have gone from 3x toward 52x. Indigo's long commentary grabbed the mechanism most people skipped: "AI is already accelerating the development of AI systems… their engineers now ship, on average, 8 times as much code per quarter as in 2021–2025… but it runs into Amdahl's Law: speeding up one link only pushes the bottleneck to the links that didn't speed up — Anthropic has already hit this: human code review has become the new bottleneck" (original post). Amdahl's Law is a basic fact of computer architecture: a system's overall speed is limited by the link that was not accelerated. He also noted that the report's proposed remedies contain almost no technical prescriptions — everything rides on coordination and governance. Another number cuts sharper: the result where autonomous agents close 97% of the research gap assumes a human has predefined a clean verifier (an automatic checker that decides whether a task succeeded); on open-ended long tasks with no verifier, the same capability drops to 21%. 97% and 21% are not two contradictory numbers. They are the two ends of the same capability curve. From this you can fix a test: first ask what the verifier is, whether it is inside or outside the loop, and whether the job is a sprint or a marathon.
On the Ground
#04 The Bottleneck Moves Down to the Materials Layer
NVIDIA's cable-free Kyber rack packs 144 GPUs into one cabinet. Its orthogonal backplane pushes the signal-integrity pressure down to copper-clad laminate (CCL, the base material of circuit boards): traditional glass fiber is no longer good enough, and the replacement — harder-to-make quartz fiber — has an immature supply chain, directly jamming the shipping cadence of the next-generation Rubin Ultra. AI's bottleneck is not algorithms. It is materials. This connects into a line with the earlier observation that the bottleneck moves down one layer every six months.
#05 Power Architecture and AI Architecture Couple for the First Time
Once racks shrink, 800VDC (800-volt direct-current power) can run through the entire data hall, and the total-cost advantage of SOFCs (solid oxide fuel cells), which output DC natively, jumps accordingly. Bloom Energy and Delta Electronics, which holds a Ceres license, are both betting on this line. The transformer is a legacy of the industrial age; the AI factory has to rebuild the entire power chain. This is the most direct footnote to the heavy-industrialization story.
#06 Sovereign AI Becomes Official Anxiety for Mid-Tier Countries
Canada released its AI for All national strategy on June 4: CA$200 billion of GDP lift over five years, 250,000 jobs created, enterprise adoption raised from 12% to 60%, with the core move being government procurement acting as a strategic anchor customer. What an anchor buyer does is turn a bet with unknown output into a cash flow you can take to financiers. From NASA and the lunar module to Anthropic and Colossus, this framework has now been hit by independent events two weeks in a row.
#07 AI in the Middle, Humans at the Edge
Modern organizations are, at bottom, middle-layer routers that forward information layer by layer — and lose information layer by layer. AI, for the first time, can orchestrate the whole board: AI coordinates from the center, while humans retreat to the edge to do intuition, trust, culture, and negotiation. Copilot mode is just a transitional state where the org chart hasn't changed yet. The middle layer is not being optimized away. It is being routed around. Human code review becoming the new bottleneck inside Anthropic is this framework's lab evidence.
#08 The Career Barbell and the Valley of Death
At one end, a high-premium boutique zone where a few people work with agents. At the other, a low-price commodity zone that is fully AI-automated. The ordinary professionals in between fall into the valley of death, and Indigo estimates this squeeze may complete within a year or two. What disappears is not the total number of jobs. It is the middle ground. This does not contradict the judgment that humans are scarce — the scarce ones were only ever the irreplaceable kind.
#09 The Real Shortage Is Deployment, Not Applications
Indigo believes what's missing right now is not another AI application but FDEs (forward-deployment engineers — the role that embeds on site and wires models into a client's real workflows). Silicon Valley dreams of changing the world with one click, while the tedious work of landing things is exactly the opportunity Chinese teams are good at and the market underrates. Deployment is not a supporting role. It is the key trade in this round of value realization. Without deployment, heavy industrialization is just capacity that never pays out.
#10 The Dark Side of Liquidity
In his Hong Kong talk he warned: if the three giants — OpenAI, SpaceX, Anthropic — go public one after another within the year, they will siphon money out of the secondary market. Hong Kong retail leverage sits at 2.8, already high, and after the semiconductor crash the leverage has not been flushed out. The rise will be slow; the flush will be violent. Crisis always arrives faster than prosperity. SpaceX's 555.6 million new shares plus a 15% overallotment are the first concrete sample of that siphon.
Slow Thinking
Indigo made one clear reversal this week, on where NVIDIA's premium actually comes from. He used to explain the premium as technical lead plus CUDA (NVIDIA's software ecosystem) lock-in. Now he has swapped in a different underlying logic — the essence of the premium is having locked up, in advance, full-chain capacity across power, land, grid connection, cooling, optical modules, memory, advanced packaging, foundry, and networking, then packaging that purchasing power for export under the AI factory concept. When Jensen Huang names a company on stage and says whose stock can rise 30%, that is this purchasing power showing up in prices. What triggered the reversal was the DSX platform launch at GTC Taipei, layered with the micro-level confirmation on the Computex floor that materials, power, and optical interconnect are all under strain across the chain. There was also a small methodological update: when reading first-party reports like Anthropic's, he moved from believe it all or discount it all
to honesty lives in the numbers, position lives in the framing
— the parts where the authors admit weakness are usually true, while what they choose to disclose, and the frame they shoot it through, is where the position actually hides.
The other side: the strongest rebuttal to the compute-as-real-estate story is whether the rent gets paid. $7.5 trillion is roughly 5% of global GDP. If the distant margins never materialize, this is not heavy industrialization — it is the largest capacity misallocation in history. SpaceX itself is the exhibit for that risk: GAAP net income of -$4.9B, with the valuation almost entirely staked on the far future. The sovereign-AI side has cracks too: Canada's strategy press release named zero dollar amounts and zero companies. The actual money is only the existing CA$2 billion range — a rounding error next to the hundred-billion-dollar capex of a single US company — so the anchor buyer may well stop at rhetoric. Watch three falsification signals: real subscription and the aftermarket after SpaceX prices on June 11; whether materials like quartz fiber ramp on schedule; and whether government procurement ever lands as concrete contracts with dollar amounts attached. If all three fall through in sequence, this issue's theme should be downgraded to a capex narrative feeding on itself.
Indigo on X
"These past few days I've been advising many friends: choosing a free place to live matters more than anything. 'Freedom of information, freedom of movement, freedom of money, freedom of expression.' Stay constrained long enough and your mind gets institutionalized too🤪"
From @indigox, 145 likes
"Anthropic Institute research shows: AI is already accelerating the development of AI systems… their engineers now ship, on average, 8 times as much code per quarter as in 2021–2025… but it runs into Amdahl's Law: speeding up one link only pushes the bottleneck to the links that didn't speed up — Anthropic has already hit this: human code review has become the new bottleneck. 'The speed of finding and fixing bottlenecks' may become an organization's most important skill… the report's 'what should we do' section has almost no technical prescriptions — it's all coordination and governance. The core bet is 'whether a verifiable global pause mechanism can be built,' and they themselves admit this is far harder to detect than nuclear weapons"
From @indigox, 33 likes
"Every time Boss Huang presents, AI Factory is the main event! We have gone from chasing GPUs to chasing 'G-megawatts.' The NVIDIA DSX platform provides a complete solution for AI factories. This is the 'heavy industry' of the intelligence era⚡️… NeoCloud is the 'silicon-based' real estate developer of the new era"
From @indigox, 30 likes
Closing
One Thought
The British railway mania of the 1840s is the nearest mirror: at its peak, railway investment swallowed an unprecedented share of Britain's national investment, and the crash of 1847 wiped out a generation of shareholders — but the track that got laid did not disappear. It reshaped the economic geography of the next hundred years. The rule of infrastructure manias seems constant: the curve is real and the price is wrong, both at the same time. Compute-as-real-estate will most likely be the same: the 150GW will get built, the $7.5 trillion will get spent, but every price paid for it will not necessarily be reasonable. In judging a mania, the question is never whether the direction is right — it is how many years of the future the current price has borrowed.
One Exercise
Spend 30 minutes doing due diligence on yourself. List roughly ten pieces of work you actually did last week. For each, ask three questions: Does it have a clear acceptance standard (a verifier)? Can a machine judge the acceptance, or does it require human judgment? Is it a sprint that shows results in hours, or a marathon that stays uncertain for months? Mark the machine-verifiable sprints — on this week's evidence, that part gets automated first. The remaining items, the ones that depend on embodied experience, trust, and human judgment, are where your time is genuinely scarce and worth investing over the next few years. The shorter the list, the more you should worry; the longer the list, the more you can relax.