This week's public signals and research materials point to the same conclusion: the drop in software stocks is not a mood swing. It is a change in the denominator used to value them. Within one week, Indigo went from leaning toward pure software is uninvestable
to stating it outright — and he also drew the boundary of the exception. That is exactly what this issue takes apart.
2026.04.26 — 2026.05.03 · Once a week: spot the signals, recalibrate.
This Week's Signals
Three sets of facts are worth recording this week. First, the sell-off in software continued. Indigo's most widely shared post of the week (290 likes) put it bluntly: "The current sell-off in software is just the appetizer! Semiconductors will eat all of software's Margin" (original post) — Margin here means gross margin, what is left of revenue after direct costs. Second, Alphabet's Q1 2026 earnings beat expectations: Cloud revenue grew 63%, operating profit was $6.6B, and the stock rose 7% after hours. Third, courts in Hangzhou and Beijing each ruled that deploying AI is a company's own strategic choice and does not count as a major change in objective circumstances
under Article 40 of China's Labor Contract Law — the first time the courts have put a brake on AI replacement.
Most people read this decline as profit-taking on the AI trade, or as one more growth-stock correction. Fewer people noticed that the denominator changed. What the market is repricing is not a handful of software companies. It is the old paradigm that being able to write software is itself a moat. When the marginal cost of writing code approaches zero, value can only move in two directions: down, into the physical layer that cannot be conjured — memory, power, packaging; and up, into the people who cannot be copied. The sell-off is the start of this value migration, not the end.
Over the next 12 to 24 months, the hard test of a software company's fate is not the growth rate of its AI ARR (annual recurring revenue from AI products). It is real deployment and retention.
Which Way the Wind Blows
#01 The Denominator of Software Valuation Has Changed
Put this week's materials together, and the heaviest sentence is Indigo's own: The software market is only 1 trillion; the labor market is 80 trillion. AI is not a continuation of the last internet wave — it is more like the age of electricity: what it will rewrite is the entire division of human labor
(original post). This explains why software is falling while compute is rising: the denominator used to value software is switching from 1 trillion to 80 trillion, and pricing power is changing hands with it. Demand-side evidence has already appeared — Google Cloud disclosed that Gemini Enterprise's token throughput (a token is the smallest unit of text a model processes, and the unit of billing) rose from 10 billion per minute in January to 16 billion per minute now, with enterprise users still growing 40% per month. Money has not left AI. It has simply bypassed the pure-software layer. Stanford's research adds a rather awkward footnote: 41% of YC's investment bets were pointed the wrong way — the automation that capital funded does not match what workers actually want. The denominator has changed. Anything that looks cheap by the old denominator is an illusion.
#02 An Extreme Amplifier: Between the Two Peaks Lies a Valley of Death
Indigo's second most shared post this week (147 likes) was not about technology but about social structure: AI is not a tool of equality — it is an extreme amplifier! The first flashpoint has already happened: youth unemployment, already stubbornly high in Canada… in between lies a valley of death
(original post). The data supports this twin-peaks picture. Stanford's WORKBank (a large paired survey of workers and AI experts across 104 occupations) shows that in 45.2% of occupations (47 of the 104), what workers most want is to work with AI as equals, not to be replaced; 46.1% of task workers view AI positively, but in arts and media the figure is only 17.1%. On the other side, Andrew Yang estimates that 70 million white-collar workers in the US face job cuts of 20% to 50%, with middle management first in line. And the courts have started to brake the will of capital: the Hangzhou and Beijing rulings mentioned above effectively write a retraining buffer into the cost formula for any company deploying AI in China. This is not policy risk. It is structural risk — it directly rewrites the market-size math for replacement-type businesses in China.
#03 The Real Bottleneck Is the Physical Layer
On the compute side, Indigo's judgment this week is that the supply-side shortage is far from fully priced in. "The three-step token survival rule… Anthropic's gross margin has surged from 30% to 72%+… DRAM prices will still double or triple" (original post). The supply-demand mismatch in DRAM (general-purpose memory chips) will, by Dylan Patel's estimate, last all the way to 2028. The physical layer's ledger is full of hard numbers: AI data centers cost about $60M/MW to build, with GPUs taking 75% of that; the Abilene, Texas campus is planned at 2.1GW, the electricity use of two Denvers; TSMC's 2028 capital spending may reach $100B, and its own technology forum points the direction — value is migrating from wafers to packaging and system integration (the process step that assembles multiple chips into one module). What is holding AI back is not imagination. It is physical supply. Once software's margin is eaten, this is the layer it goes to.
On the Ground
#04 The Ceiling on Context Is the Memory Wall
What is holding back model context length? Not compute, but the memory bandwidth bottleneck
(original post). First-principles math shows that 100K-200K context runs straight into the memory wall: the KV cache (the memory the model uses to hold prior text while generating) costs about 2KB per token — which is also the micro-level reason the memory price rises in #03 will persist.
#05 The Overlooked Second Shortage: CPUs
Everything around the CPU is still scarce, and it will gradually split off into its own logic
(original post). The market confirmation: CPUs are completely sold out. While every eye is on GPUs, general-purpose compute is quietly becoming the physical layer's second bottleneck.
#06 Training Sets the Ceiling, Inference Sets the Floor
On the same GPU, the efficiency gap between different inference systems (the computing architecture that serves a model after launch) can stretch to 14-105x. Anthropic and Amazon have locked in 5GW of compute; Meta's contract with Nebius runs up to $27 billion over 5 years. Inference is the part of the physical layer closest to cash flow — the tollbooth where the 80 trillion denominator actually gets collected.
#07 The Grid Hits the Wall First
The December 2025 capacity auction in PJM, America's largest grid operator, exposed a 6.6GW shortfall, and transformer lead times have stretched to 128 weeks. AI's next bottleneck is not chips. It is electricity — the physical-layer opportunity now extends from silicon to grid equipment.
#08 Models Are Commodities; the Harness Is the Moat
In this week's materials, the line models are commodities; the harness is the moat
appeared independently in 6 places — Karpathy, Notion, CompanyOS among them. It does not contradict #01: what is depreciating is generic software; what is appreciating is the layer bound to workflows and proprietary data.
#09 The Most Expensive Thing Will Be People
The AI era may bring a… reversal: in the future the most precious thing will not be having AI, but having humans. Human teachers, personal trainers, and companionship will become luxuries
(original post). Among WORKBank's 104 occupations, editors are the only one whose workers want no AI at all — people that scarce are exactly the residents of the left peak in #02's twin-peaks picture.
#10 The Gateway to Our Time Has Changed
Anyway, I already spend more time with Agents than on social networks
(original post). Agents (AI programs that carry out multi-step tasks on their own) are replacing the feed as the new gateway to our time — this personal-level shift in media is the smallest sample of the 80 trillion rewrite of human labor.
Slow Thinking
Indigo made one clear upgrade to his position this week. Before, his framework followed the chain electricity → chips → compute → models → applications,
favoring the shovel-selling positions while keeping a cautious sliver of exposure to application-layer software. Now he has turned that vague preference into a public hard filter: Pure software is uninvestable! If your only edge is writing software others cannot write, that moat is now gone
(original post). Three things triggered it: Naval put forward the same thesis; Karpathy showed a complete application, MenuGen, being replaced outright by one prompt (an instruction given to a model) plus Gemini; and data from the inference cloud Baseten — more than 95% of the inference tokens on its platform come from post-trained models (general models further trained on a company's own data), and its NDR (net dollar retention, what existing customers spend next year versus last year) runs as high as 400%. The same data also draws the boundary of the exception to the hard filter: the only application-layer companies still worth watching are closed-loop businesses that hold exclusive user signals and can turn those signals into training feedback for a model.
On the other side, the strongest counter-argument is that the speed of replacement is being systematically overestimated. Worker preferences (45.2% of occupations want to work with AI as equals), the legal brake (the two Chinese rulings force companies to add a retraining buffer to AI deployment), and evidence that a single occupation spans multiple replacement quadrants all suggest that this occupation will be wholly replaced
is close to a false proposition. If replacement moves slowly, software companies have time to fold AI into their existing distribution channels, and the old moats may not actually disappear. The falsification signals are also clear: if over the next 12 to 24 months a group of pure SaaS companies show consistently strengthening real deployment and retention with gross margins not eroded by compute costs, or if DRAM price increases ease visibly before 2028, then this issue's software valuation collapse
thesis has to be rewritten.
Indigo on X
"The current sell-off in software is just the appetizer! Semiconductors will eat all of software's Margin"
From @indigox, 290 likes
AI is not a tool of equality — it is an extreme amplifier! The first flashpoint has already happened: youth unemployment, already stubbornly high in Canada… in between lies a valley of death
From @indigox, 147 likes
What is holding back model context length? Not compute, but the memory bandwidth bottleneck
From @indigox, 124 likes
Closing
One Thought
In the years when electricity replaced steam, the early winners looked like the makers of generating equipment. But the real redistribution of wealth happened on the factory floor: a steam-age factory was laid out around one central drive shaft, while the electric motor let every machine run on its own — so the whole production line, and the whole management hierarchy, got redrawn. The companies selling electricity did not take most of the profit; the companies that reorganized production around electricity did. Mapped onto this week: the model is the power plant, tokens are the electricity, the harness is the electric motor — and the position that produced the biggest winners last time was always the one that reorganized the division of labor, not the one that generated the power.
One Exercise
Take 30 minutes. Pick the software company you know best (or just use your own job) and run it through the five screening questions from that inference-cloud founder this week: One — is its core capability a generic model with an interface on top, or a post-trained specialized model? Two — does it have user signals nobody else can get? Three — can those signals become feedback that trains a model? Four — do customers spend more in year two than in year one, or less? Five — has it arranged anything on the compute supply side? Write down the answers, then reread #01 — you will know quickly whether it stands on the 1 trillion side or the 80 trillion side.