Two years ago, Zaharia and colleagues at Berkeley named a shift that was already happening: state-of-the-art AI results were no longer coming from monolithic models, but from compound systems with multiple interacting components. The thesis has aged remarkably well. Every frontier deployment of consequence in 2025 and 2026 is a compound system. Retrieval, tool use, agentic loops, chains, ensembles, routers, test-time compute scaling, speculative decoding, mixture-of-experts dispatching. These now define how the best AI products achieve quality, controllability, and economics that no single model call can match.

Its hardware substrate has not.

We argue that the compound systems thesis has a silicon layer counterpart that is no longer optional. State-of-the-art AI deployments increasingly run on heterogeneous hardware, and the orchestration layer above that hardware is becoming the locus of competitive advantage in the same way the orchestration layer above models did at the application stratum. Call it the shift from compound AI systems to compound compute substrates.

What a compound compute substrate is

A compound compute substrate is the execution surface of a single workload whose operations are assigned across two or more architecturally distinct classes of silicon, where that assignment is a design variable rather than an artifact of what happened to be available.

Three things in that sentence are load-bearing.

Single workload, not fleet. The parallel to Berkeley’s thesis only holds at the level of one job. A compound AI system is one task spanning several model calls, retrievers, and tools, not a datacenter that happens to contain several models. The substrate version is one inference, or one control loop, spanning several kinds of chip with coordination measured in microseconds.

Architecturally distinct needs a functional test, not a taxonomy, or the term dissolves into vendor marketing categories. Two substrates are distinct when an operation well mapped to one runs at a small fraction of achievable efficiency on the other. That keeps a CPU and a systolic-array ASIC apart, keeps two generations of the same GPU family together, and lets the boundary move as workloads change.

Assignment is a design variable. This is the clause that matters most and it is the one an earlier draft of this piece got wrong. The definition does not require an orchestrator to exist. The assignment can be made by a person, expensively, and today it usually is. Compound compute substrates are already in production. What is missing is a system that makes the assignment automatically, which is a claim about tooling rather than about deployments.

The taxonomy falls out. A fleet of one silicon class is homogeneous, not compound. A mixed fleet where every job runs entirely inside one class is heterogeneous procurement, still not compound. Only a workload that spans classes is compound. And the Mismatch Tax, defined below, is what you pay when a workload that wants a compound substrate gets executed on a uniform one instead.

The reason to want one is that a modern AI deployment is not a single computational primitive. A representative production pipeline includes dense matrix multiplication, sparse computation, mixture-of-experts routing, attention over variable sequence lengths, key-value cache management, retrieval over vector indices, tokenization, sampling, speculative decoding, draft model verification, encoders across several modalities, tool invocation, structured output parsing, and a substantial amount of CPU-bound control logic. Each has a different optimal silicon profile.

Vendors do not converge on one architecture because the physics will not let them. A GPU’s strength is throughput on dense compute; its weakness is control flow. A TPU’s strength is matmul performance per watt; its weakness is anything outside the systolic-array design point. An NPU sips power on quantized inference and chokes on large unstructured contexts. Nobody has built the chip that optimizes simultaneously for throughput, control, latency, memory bandwidth, power efficiency, and reliability. Nobody will. It is a physics gap, not an engineering gap.

Why use them

The empirical case starts with a number, and the number has a name attached.

Anjney Midha, founder of AMP PBC and previously the general partner who allocated a16z’s GPU cluster, gave the most specific public accounting on TBPN on 23 July 2026. His decomposition has two layers. At the fleet layer, a lab taking a long-term lease loses 30 to 40 percent of its FLOPs to scheduling alone; nodes sit unallocated. At the chip layer, model FLOPs utilization runs around 15 percent, because the chip waits on storage, memory, networking, or a handoff. Together he puts realized utilization at roughly 15 percent of every leased dollar, and calls the remaining 85 percent “a national security crisis.”

Both layers are measured, and it is worth asking why. Allocation is logged, so scheduling loss is visible. Vendors publish peak FLOPs and labs report realized ones, so utilization is a ratio of two disclosed quantities. In both cases the instrument existed before anyone wanted the measurement.

There is a third layer and it has no instrument.

Scheduling loss is capacity that was never allocated. Utilization loss is a chip waiting on its own memory hierarchy. The third thing is different in kind: the chip is allocated, it is busy, it is running at its own design point, and the design point is wrong for the operation in front of it. Call it the Mismatch Tax.

The three do not add. Scheduling and utilization loss are close to multiplicative. Mismatch is not a third multiplier; it is a reweighting of what the surviving FLOPs are worth. A dense matrix engine running control logic at a few percent of CPU efficiency looks fully occupied to every available instrument while producing almost nothing. So mismatch lives partly inside the 15 percent counted as realized and partly inside the 85 percent counted as lost, and no partition of the two recovers it.

Which means the term names a mechanism, not a rate. That is deliberate, and it has a precedent: Hooker never quantified the hardware lottery either, and said why.

Key challenges

Design space. For each workload class the orchestrator must decide which substrate family executes it, at what precision, with what memory placement, on what schedule, and how data is staged when tensors cross family boundaries. Twenty workload classes and six substrate families is already a combinatorial explosion, before dynamic re-routing under thermal pressure or workload-mix shift. A policy optimal at peak load is suboptimal off-peak. A policy optimal for the current model architecture is suboptimal after the next refresh.

Optimization. This is the silicon-layer analog of the problem DSPy addresses at the model layer, and the correspondence is close: both action spaces are discrete plus continuous, both cost models are empirical rather than analytic, both feedback signals are end-to-end rather than gradient-based. The substrate version is harder on two axes. The action space is larger, and the feedback signal costs wall-clock plus joules on physical hardware rather than task accuracy on a held-out set. Eighteen years of work from LATE onward establishes that learned scheduling over heterogeneous resources beats hand-tuning; the problem extends that lineage rather than starting from scratch.

Operation. Thermal throttling, memory pressure, congestion across substrate boundaries, and workload-mix shift all perturb realized cost. The orchestrator must monitor telemetry and re-route in milliseconds.

Measurement. This is the challenge the model layer never had. Isolating the third loss layer requires running one workload mix across architecturally distinct fleets at matched power and comparing. The parties who can run that comparison are the hyperscalers and the chip vendors, and neither has a reason to publish a number saying their silicon is the wrong silicon for a large share of the work. The measurement is not hard. It is unincentivized, which is worse.

The recursion

Here is the part most readers miss on first encounter: the compound AI systems thesis is itself the mechanism making the substrate problem worse.

Every component added to a pipeline brings its own optimal silicon. Vector retrieval is memory-bandwidth-bound. A symbolic verifier is control-heavy. A speculative decoding draft model wants a different precision than the verifier it works against. A tool-calling loop interleaves control flow and inference in a pattern no single substrate runs well.

Write it as a model. A compound system with k components, a probability p that the fleet provides the right substrate for any one of them, and a mean penalty m when it does not. Aggregate tax is approximately k × (1 − p) × m.

All three terms move the wrong way. Component counts rise, because each generation is more composite than the last. Match probability falls, because the substrate landscape keeps diversifying. Per-component penalty rises, because specialized silicon keeps widening its advantage on design-point workloads.

The product compounds rather than accumulates.

The rates are illustrative. What survives without them is the sign of each term, and each sign is checkable. The recursion is wrong if pipelines consolidate back toward single model calls, if the substrate landscape consolidates rather than diversifies, or if specialized silicon stops pulling away from general silicon. Anyone who believes one of those is happening should discard the argument.

The asymmetry is the point. The more successful compound AI becomes at the application layer, the more expensive the missing layer underneath it gets.

Emerging paradigms, and one that was removed

Four research threads are converging on substrate orchestration. Heterogeneous scheduling, from LATE through FlexFlow, POP, Pathways, and Alpa, establishes the optimization is tractable. Inference economics, from FrugalGPT through vLLM and RouteLLM, establishes that learned routing pays. Compound AI compilation, with DSPy as the canonical case, establishes the framework. Substrate-aware compute accounting establishes the governance case, since a FLOP threshold meaningful on a uniform fleet is not meaningful on a mixed one.

The fifth thread is the existence proof, and it is missing. Not because nobody tried.

OctoML, spun out of the Apache TVM project in 2019, built substantially this system: a compiler and serving stack routing models across CPUs, GPUs, and accelerators including AWS Inferentia. NVIDIA acquired it, as OctoAI, in September 2024 for a reported 165 million dollars against a 2021 valuation near 900 million. Hosted services terminated on 31 October 2024, five weeks later, with no successor. Red Hat acquired Neural Magic in November 2024, whose vLLM work spans AMD GPUs, AWS Neuron, Google TPUs, Intel Gaudi, NVIDIA GPUs, and x86. NVIDIA had already taken Run:ai, and later CentML.

The vendor-neutral layer has been built several times. Each time, an incumbent bought it.

Two readings. The first is that the category failed on demand: buyers with real hardware diversity were large firms with internal teams, and the long tail was on NVIDIA through the clouds anyway, so the acquisitions were rational exits from a thin market. The second is that the record confirms the incentive argument rather than refuting it: a neutral routing layer is worth more to the dominant vendor as a removed option than to the market as a product, and low-hundreds-of-millions prices followed by a five-week shutdown is not what buying a business looks like.

The distinguishing variable is substrate diversity on the demand side, which is exactly what the recursion above predicts will grow. The acquisitions are evidence about the size of the 2024 market. They are not evidence about the 2028 one.

What this means

For AI developers, the substrate question is load-bearing. Choosing between vendor stacks is no longer the right framing. The right framing is choosing an orchestration layer that lets a workload run on whichever substrates serve it best while preserving the option to integrate new ones.

For infrastructure operators, the Mismatch Tax will not be closed by kernel optimization, by better scheduling, or by waiting for the next-generation GPU. Kernel work improves the implementation of an operation on a substrate. Scheduling allocates nodes. Neither chooses between kinds of node.

For policy, substrate-aware compute accounting is the necessary instrument for any quantitative regulation of frontier compute. The EU AI Act presumes systemic risk above 10^25 FLOP; the US 10^26 threshold outlived the executive order that created it and now sits in California’s frontier AI act. A quantity surviving the repeal of the instrument that introduced it is decent evidence the underlying assumption is entrenched.

Three calls

By 31 December 2027, custom AI accelerator silicon exceeds 25 percent of combined datacenter AI accelerator revenue, computed from disclosed filings as Broadcom AI semiconductor plus Marvell custom silicon over that sum plus NVIDIA data center plus AMD data center GPU, trailing four quarters. Wrong if the ratio is at or below 25 percent. This adjudicates the two readings of the acquisition record.

By 31 December 2028, no major chip vendor or hyperscaler ships a generally available orchestration product that routes production inference to a competitor’s silicon by default. Wrong if NVIDIA, AMD, Google, AWS, or Microsoft reaches general availability on a product whose documented default includes routing to a named competitor’s accelerator.

By 31 December 2028, at least one published like-for-like comparison of a mixed frontier workload on homogeneous versus heterogeneous fleets at matched power reports a gap above 25 percent. Wrong if nobody publishes, or every published comparison lands at or below 25 percent. This is the one that loses if the central claim here is false, and it is built so that silence counts against the thesis rather than for it.

The compound AI systems thesis told developers the locus of competitive advantage at the application layer was moving from models to orchestration. The compound compute substrates thesis says the same shift is now happening at the infrastructure layer, for the same reasons and with the same mathematics.

The Mismatch Tax is structural, it is unmeasured in public, and the layer that would both close it and measure it does not exist.

This essay extends the framework Zaharia and colleagues introduced in The Shift from Models to Compound AI Systems (BAIR Blog, February 2024). The full treatment, including the historical record of substrate orchestration attempts and the argument that the term names a mechanism rather than a rate, is in the working paper.

Stanley,

There is a job at Amazon called Amnesty Floor Monitor.

Amnesty is what Amazon calls an item that ends up loose on the floor of a robotic field, blown there by the airflow of a few hundred drive units in motion or pushed out of a pod by an arm that failed its stow. The Amnesty Floor Monitor keeps the floors clear and resets units when needed. That is the whole role. A person, in a safety vest whose green light tells nearby drives to hold, walking a caged grid, picking up what the automation dropped.

I know that role well, because I was there when we invented it. In 2012 we called it Flow Assistant.

Amazon was a customer of Kiva before it became Kiva’s owner. I was at Amazon on the team that ran those initial deployments, from 2011, through the acquisition in March 2012, and into the build of the first fully autonomous fulfillment center. No one’s design docs included Flow Assistant. We found we needed it in the first weeks of live operation, because inventory came off pods and ended up on the floor. And a drive unit that meets a fallen item is a drive unit that stops, and a drive unit that stops in a dense grid is the first bead in a jam that propagates. So we put a person down on the floor. This was meant to be a temporary thing.

Fourteen years later, it has an official title, a line for headcount, and a published failure rate.

In a paper covering March 2025, Amazon’s robotic stow team reported that roughly 3.7% of stow attempts pushed an item onto the floor. Stow success was about 85% across more than 500,000 attempts. The companion pick robot succeeded on 91% of what it attempted but declined 19.4% of the requests it was handed. On the same floor, in the same month, human associates stowed 243 units an hour against the robot’s 224.

Those numbers are from the most advanced applied physical AI deployment on earth, thirteen years and a million units in. They refer to a system that is slower than a person, rejects one job in five and spews out a steady stream of debris that someone is paid to clear up.

That system is a victory. I’m not being sarcastic. It is the right model to study. But here is the finding I want to present to you, and it’s why I’m writing instead of watching:

The recovery role was not removed. Not more than fourteen years. Six generations of hardware. A thousandfold increase in fleet size. A company with functionally unlimited capital and the best applied robotics team in commerce. It was renamed, regularized, staffed and metered. It was never engineered out.

Your team will have someone who will tell you that your equivalent of the Flow Assistant is transitional, scaffolding that you retire when the stack matures. We thought that in 2012. And we were wrong. And we have been wrong for fourteen years.

What you actually bought on wednesday

First, the congratulations, and it is specific. DoorDash is the eighth drone operator to earn Part 135 air carrier certification in more than a decade of attempts. That’s an organizational achievement, not a hardware one. Someone wrote an operations manual, developed a maintenance program, brought in a director of operations and chief inspector, established a records system and passed a five-step FAA evaluation. Most engineering organizations do not produce that set of documents. That’s why there are eight names on the list.

Now for the uncomfortable part. A Part 135 certificate is not a permission slip. It’s a standing obligation whose maintenance cost scales with the number of different operational configurations you run, and not at all with the number of deliveries you make. Every new aircraft type, base and material procedure change nibbles at the scarcest resource you have now - the attention of the few people who can hold your safety case in their heads and defend it to a regulator.

You didn’t buy scale. You bought an obligation and moved your binding constraint to a place where you have less experience running it.

The arithmetic under your own diagnosis

On No Priors you said the software stack is no longer the primary barrier, that the real ones are operational integration and hardware manufacturing, and you gave the sharpest possible example: fleet boot scripts that work for ten robots and break at five hundred.

That’s not a code quality problem. p is the probability that a unit will not be ready to go on any given morning, for any reason. Stale certificate, DHCP lease, timeout in the merchant portal, firmware skew, charger that said full but wasn’t... At ten units and p = 2% an exception occurs every fifth morning and one person takes it with a phone. At five hundred units and the same p, ten exceptions come every morning, before the breakfast rush, across a metro. You can’t get that on a phone.

The per-unit reliability required for constant human attention scales with fleet size. The operator’s morning is still good. Five to ten hundred. P needs to be fifty times better. Nothing scripting gives fifty times.

Amazon ran the natural experiment. The fleet went from roughly 200,000 drive units in 2019 to over a million in July 2025 across 300-plus facilities. If robots substitute for the humans who tend them, support headcount per facility should have fallen.

It rose.

Amazon says its new generation of fulfillment centers need 30% more workers in reliability, maintenance and engineering jobs. For sourcing hygiene: 25% according to Tye Brady at TechCrunch, 30% according to Amazon’s written announcement. I cannot reconcile them, nor will I deny that there is a discrepancy. In either case the sign is the same, and the sign is the finding.

So Amazon didn’t get a p low enough to pull the human out of the recovery path. It did two things at once. It drove p down hard, and built a permanent human organisation around the residual.

Its leaders tell the truth about this. “If we had to have Vulcan do 100% of the stows and picks, it would never happen,” says Aaron Parness, who leads the science behind Vulcan, Amazon’s first robot with tactile sensing. In that framing, the 19.4% pick-rejection rate is no defect. It is a work of design. The system knows what it is to avoid.

The two amazon programs that are already yours

Scout is Dot. Prime Air is Air. Amazon ran both of your programs, in the order you are running them.

Scout launched in January 2019, a cooler-sized six-wheeled sidewalk robot, expanding to Irvine, Atlanta, and Franklin, Tennessee. Each unit was accompanied by a human Scout Ambassador. In October 2022 Amazon ended field tests and disbanded the team of roughly 400 people, citing aspects of the program that were not meeting customers’ needs.

See what didn’t work and what did. The robot did its job. For nearly four years it drove sidewalks in four climates. What didn’t work was the economics of a machine that required a person to walk beside it.

Dot isn’t a Scout. It runs roads and bike lanes at 20 mph instead of crawling a sidewalk, it carries six pizza boxes, and it is dispatched by a network that already knows what a good handoff looks like from billions of deliveries. These are the proper differences. But Scout is the precedent that a robot that can’t do the last ten feet without a companion is capability complete and network incomplete, and the gap between the two states can absorb four years and four hundred people before anyone calls it.

The question is not whether Dot can drive. It certainly can. What is your ratio of human interventions to completed deliveries, what is its trend across successive waves of deployment and below what number do the unit economics close? It is: If no one can answer the third part, then the program doesn’t have a definition of done yet.

Prime Air has been in operation since 2013. As of February 2026, it had made about 16,000 lifetime deliveries, nowhere near its stated target of 500 million a year by the end of this decade. “Between one and thirty thousand somewhere.

What is useful is the history between those numbers. There were at least eight test crashes in a twelve-month window in the MK27 era, one starting a brush fire. Mid-air crash during engine failure simulation. Two crashes in Pendleton December 2024. Two-month service break. Resuming after dust in dry Arizona air, a condition Texas never had, impacted altitude sensing fixes. In October 2025, two aircraft struck the same construction crane in Tolleson within minutes of each other. In November, one severed a residential internet cable in Waco thirteen days after launch. In February 2026, one hit an apartment building in Richardson.

The validation distribution did not have environmental or temporal domain shift. Most of these are one failure in different clothing: A regime of precipitation. A system of particles. Same-day change in the vertical obstacle geometry.

The crane Tolleson wants is the one to print out. There is no aircraft failure at all. Two aircraft encountered the same new problem minutes apart, with the second not knowing what the first had just learned. That is a breakdown of fleet knowledge. You have one vehicle, and it does not disappear by making the vehicle better. It does not exist.

Two asks come out of that decade. Propagate discovered obstacles fleet-wide, in minutes, not release cycles. And fund environmental domain coverage as a separate budget line from behavioral coverage, because engineers naturally enumerate the construction zone and the off-leash dog, while environmental regimes are what actually ground fleets and are boring enough that nobody volunteers to own them.

The argument for the position you already hold

You said the future is more Dashers, not less. I think more properly than you put it.

Weather degrades air capacity in an ungraceful manner. It turns it off, simultaneously, spatially. And in the worst possible way, the covariance. Rain grounds aircraft, rain reduces Dasher supply because fewer people want to drive in the rain, rain raises demand because fewer people want to leave the house. At the point where the air drops to zero, ground fallback is thinnest and demand highest.

The only mode of which availability is imperfectly correlated with all others is human supply. That is therefore the reserve asset that prices the risk in each autonomous mode. If anything, that argument only gets stronger as autonomy scales, and the proof is Amazon’s history: a million robots in and the human organization around them grew.

In practice: price reserve against joint distribution of mode availability, Dasher supply and demand. In code, make sure no order goes to a mode whose abort has no ground fallback inside the food-quality window. Weather is not an availability haircut. Don’t model it that way. A haircut is an independence assumption, and independence is the one thing weather is not.

One question for your leadership team

In the footage you put out on Wednesday there is a shot of the workshop. Benches, parts bins, cable spools, an airframe in pieces, a rotor spinning inside a test cage. I recognize that room. Ours looked like it in 2011, and I mean that as a compliment, because it is the correct room for this stage. The building with LABS painted on the roof is your sandbox.

We didn’t grow Amazon by adding robots. We scaled through proving a facility. Sandbox, template, replicate, with discipline living entirely in the freeze. A template that is changing all the time is a permanent sandbox with a bigger budget, and the replication economics never come. Thirteen years later, the Amazon unit of scaling is still the building, multiplied across some forty next generation sites by the end of 2027.

The sandbox is not the achievement. The specification that leaves it is, along with whether anyone is allowed to keep changing it afterward. Which is the question:

What’s your template unit? And have you frozen it?”

Conveniently there is a facility with walls and so for Amazon it is a facility. Your network doesn’t have one so the template unit is a design choice, not a physical fact, and the wrong choice is invisible for two years. A metro is the obvious answer and I think it is wrong, because metros are different in ways that you cannot control. Site archetypes make better candidates: the drive-through, the strip-mall back door, the rooftop, the apartment lobby, the suburban driveway. Those are repeated in metros. Freeze a handful, prove the handoff for each one. Then a new metro is a composition problem, not a discovery problem.

One more property, as it determines how carefully you choose the first ones. The sandbox is not permanent. The site where a system is first made to work accrues operating hours and institutional knowledge that no other site can buy, and it tends to stay among the strongest in the network long after everyone stops calling it the experiment. Pick your first archetype as if you’ll still be running it in 2036, because you probably will.

The one thing to do

Instrument the recovery layer now while it is a line item not a crisis and give it an owner senior enough to be inconvenient.

Two numbers, published weekly, at same seniority as delivery volume :

a) Human minutes per day of unit cause based breakdown. Your real cost of goods. And the real test of whether these assets are autonomous.

b) Interventions per finished delivery and its trend over deployment waves. If that number does not go down as you scale, you are still capability-limited, thinking you are network-limited.

Neither will be important until the quarter where it is the only thing people are talking about. And that is exactly why it needs a person and not a working group. DoorDash Labs is a startup inside an enterprise, which is the right structure, the same structure that produced everything Amazon has in robotics, but it has one characteristic failure mode: a startup inside an enterprise is measured on shipping, and recovery infrastructure never ships. No announcement, no demo, no release date. It does not drop arguments. It never gets into anybody’s objectives, so it never gets into anybody.

Claims you can grade me on

By the end of 2027, the fully loaded cost per successful delivery for DoorDash Air will match the equivalent three-to-five-mile Dasher delivery. The crossover point is regulatory, not technological. As of 10 July 2026 Part 108 was still at OIRA. Confidence 75 per cent.

In terms of Air throughput, your first three metros will be constrained by ground handling and pad service time first and foremost, not aircraft availability or autonomy capability. Confidence 80%

DoorDash Air is expected to have at least one voluntary or regulator-prompted operational pause of seven days or longer before the end of 2028. Not a critique. A baseline rate. Confidence 65%

Support and recovery personnel per active autonomous unit will not decrease monotonically through 2028. The direct transfer of the Amazon finding, and the one I most hope I am wrong about. Confidence 70%.

No US autonomous delivery operator will publicly disclose an autonomous fault resolution rate or a human-minutes-per-unit figure before the end of 2028. The operational layer gets named freely in interviews and disclosed almost nowhere, which is itself the finding. Confidence 85%.

The most famous technical letter ever sent to an American president concluded that such bombs might prove “too heavy for transportation by air.” It was wrong, and it looked wrong after the fact, because the author put the assumption on the page instead of burying it. That is the bar. Rate me.

What is at stake

The prize is not a drone program. It’s the stance your announcement quietly makes: the mode-agnostic clearing layer for physical local commerce. When the aircraft does, or the sidewalk robot does, or the driving model does, whoever can decide in under a hundred milliseconds what physical mode moves an object for the last three miles has something that doesn’t commoditize.

Inexpensive airframes. People building autonomy stacks will find them getting cheap faster than they expect. What’s scarce is the stuff that intelligence can’t make, that took ten years to build. The certificate. The safety case. The pad. The merchant relationship. The demand forecast. The trust of a city council. The technician who can turn an aircraft around in eleven minutes. The plane is the denominator. The numerator is the ground.

“You put it in a frame.” The plane is not the whole ball of wax. That is the airline. That’s the harder work, the less glamorous work, and the work your current company is better positioned to do than any specialist you now compete with. That asymmetry is real and it won’t be permanent. I would use it.

I may be wrong on some of this. You have never seen my internal numbers, and I am reasoning from the public record and lessons bought at another company, another decade, on a warehouse floor rather than in the air. If the number of interventions is already on a downward trend, I have written you a long letter about a problem you solved last quarter and I would rather hear that than be right .

But if any of it does hit, the boring one is the thing to do. Not the airplane.

Go find out what you are already calling your Flow Assistant. Not so you can hire more of them. So you know how much of the work still needs a person to walk over to the machine, and can start designing that walk out.

Again, Congratulations. The certificate is a very hard thing, honestly earned.

Software ate the world. AI ate software. Hardware is eating AI.

The software industry is in the midst of a panic attack, fifteen years after Marc Andreessen declared that software was eating the world. Software ETFs are down 30% since January, wiping out all the gains since the launch of ChatGPT. They’re calling it the SaaSpocalypse. Everyone is asking if it’s a bubble. What matters is whether the thing being bought can be switched on.

The argument is not that AI is overhyped. AI is a physical industry priced like a software industry and the correction is not going to be a sentiment crash. It arrives in a queue. So…It’s fair to say that hardware, for the time being is eating AI.

Look at the pricing of the model layer and what is striking is not that it is expensive. That the diffusion in it has nothing to do with quality of model. At $852 billion, or about 34 times its $25 billion run rate, OpenAI is losing about $1.22 for every dollar it earns. Anthropic beat it in May at $965 billion on a bigger revenue base, a lower multiple and a first operating profit of $559 million. Train frontier models. One makes money. The difference between them is not intelligence, but the cost of servicing it. Anthropic’s gross margin improved from about negative 94 percent in 2024 to about 60 percent on inference efficiency, not pricing. The physical layer already determines who is profitable at the model layer. This is not a bubble. It is a category error about where the scarcity is.

Not all atoms are the same. And the discipline here is not to treat them as one thing. Chips, memory, packaging, steel and electricity are not all tied at the same time. The task is to determine which is the tightest.

Why is this happening now?

Not because hardware became easier, cause everything else ran out.

The software abstractions got so cheap, and so universal, that they were no longer the constraint. Frontier-level intelligence cost about $60 per million tokens at the launch of the GPT-3 API in 2020. That’s about $0.10 today: a 600x increase in six years and still climbing.

That number did not disappear. It moved to whatever the scarcity left in plenty. The Residual Ratio is how we track migration – complements over intelligence, set to 1.00 in 2020 and now near 987. The finding is the decomposition. Intelligence collapsing is about 600-fold of it. Complements appreciating is only 1.65-fold.

The physical side didn’t move. Bain expects about 200 GW of incremental compute demand by 2030, the IEA has data-center electricity rising from 415 TWh to 945, and hyperscaler capex guidance is near $725B for 2026 alone. Against that, Berkeley Lab’s Queued Up data shows more than 2,060 GW seeking interconnection at the end of 2025, with a median wait from request to switch-on of more than five years.

In February 2026, Elon Musk put it bluntly: those who have lived in software land are about to get “a hard lesson in hardware.” His example was not the chip. It was the interconnection. You go to a utility at scale, and a year later they come back to you with a study. No link. A study.

^{Musk, Elon. Interview by Dwarkesh Patel. Dwarkesh Podcast, 5 February 2026.}

So here’s the arithmetic. Intelligence falls ten-fold a year. But the median interconnection needs more than five. Electrical output outside China is flat, but accelerator output is rising almost exponentially. The conclusion is obvious and the industry’s hardest sentence: chips pile up and nobody can turn ‘em on.

^{Intelligence falls five orders of magnitude before the median interconnection is energized.}

Andreessen walked through industries, pointing out that each had quietly become a software company. This works the same way, in reverse.

The frontier lab thinks it’s on top. It leases. It calculates what it does not own, in buildings it has not built, using power it has not bought. Amazon led the March deal with a $50 billion stake, and Nvidia and SoftBank each pledged $30 billion. When your compute vendor capitalizes you, you are not the customer. You are the conduit.

The neocloud believes it owns compute. It has a depreciation schedule. It leases a sold out for two years TSMC wafer in CoWoS packaging, stacked with memory from SK Hynix, Samsung or Micron.

The hyperscaler sees itself as a tech company. It’s a utility customer with a nine-figure interconnection problem and a transformer book measured in years. It can pay any sum for electricity and still not get any in 2027 because what it needs is not electrons. It’s permission.

The chip designer believes he has the chokepoint. Nvidia has a blueprint. That design is produced by TSMC, on machines that only ASML produces, in packaging already sold, stacked with memory produced by three companies that control virtually the whole market. “Each of them is a veto. Nvidia has none of them.

The data center believes it is real estate. It’s a thermal problem attached to a substation, waiting for a transformer, waiting for electrical steel, waiting for a mill nobody built because for twenty years nobody needed one.

^{Each layer and the thing it cannot proceed without. The rust spine marks a veto capital cannot buy through.}

Everyone thinks it is the peak and everyone is waiting for someone who does not respond quickly to email. Imagine you’re a hardware company and you haven’t hit your constraint yet.

Liebig, not Schumpeter

Schumpeter was Andreessen’s call. Creative destruction was perfect for software, because the destroyer was faster than the incumbent. Not here. It’s the slower thing than AI. It’s the thing that eats AI .

The right lens is Liebig’s Law of the Minimum: yield is limited by the scarcest input, not the average. Twice the accelerators . Three times the parameters . Half the price of inference . If the substation is not energized you have generated zero additional tokens .

Many writings combine two questions. What’s stopping you from deploying? Power. Where’s the spread (margin)? Memory and advanced packaging, the layers that capital is unable to replicate in its own horizon. Power limits capacity. Memory collects the rent. Treat them as one and you will buy the constraint that is loosening as the one that is printing money tightens.

This is why “we will just build more” is not a rebuttal. Capital is fast; permits are not. A fab takes three to five years, the median interconnection more than five, a transformer order years, and verified data does not regenerate once the public web is exhausted. Money arrives in a quarter. None of these do.

And here is the part that stings. A software career trains one reflex: hit a wall, route around it. Add a layer, cache it, shard it, ship Friday. Against a physical constraint that reflex is not merely useless. It is harmful, because it produces the confident belief that a solution exists at your current velocity, and it burns the eighteen months you should have spent standing in the queue.

The hard lesson

The constraint does not negotiate. You cannot ship past a five-year queue and you cannot hire past it. Everyone faces the same physics, so the differentiator is not cleverness but position: who stood in which line, how early. Most positions that matter in 2029 are being taken this quarter, by people not posting about it.

Most of “hardware” gets eaten too. One question sorts every layer: can capital overbuild it inside its own investment horizon? Merchant generation, GPU rental and undifferentiated silicon, yes, and they commoditize. Leading-edge fabrication, advanced packaging, high-bandwidth memory, interconnection, firm power and verified data, no. “Long infrastructure” is now as coarse and as dangerous as “long software” was in 2021.

The margins already say which is which. SK Hynix reported a 58.4 percent operating margin in Q4 2025, passing TSMC for the first time in seven years, with Samsung on the same path and gross margins forecast at 63 to 67 percent against TSMC’s 59 to 61. Memory went above the leading foundry in the same quarter the consensus called compute a commodity. Commodities do not do that. And SK Hynix has said HBM4 demand exceeds its capacity for the next three years. That is what a layer capital cannot overbuild sounds like when the supplier says it aloud.

Owning the physical layer is not the end state. In 1982 Alan Kay said that “People who are really serious about software should make their own hardware.” Jobs quoted him introducing the iPhone. Run it backwards and you get the operative sentence for the next ten years: anyone serious about hardware must make their own software. A memory maker that ships parts is a supplier; one that co-designs the hierarchy against the architecture running on it owns a design space. A power developer that sells electrons is a merchant; one that co-designs against its own grid position owns the interconnection.

Call the clock on that loop loop time: from identifying a binding constraint to shipping a co-designed system that relaxes it. It runs in years, dominated by permitting rather than tapeout, which is the real reason the physical layer is winning. It also dates the expiry, and the gate is already moving: FERC's 2023 reforms landed the following year and the queue fell 10 percent in 2025. Owning a bottleneck is not buy and hold. It is buy and watch.

The bill

Building an AI company in 2026 is difficult in an older, less forgiving way: the way a refinery is difficult, or a rail line. Feedback takes years. Mistakes are not reversible in a sprint. The people who are good at it get no attention, because nobody makes a thumbnail out of a substation.

So the right posture is neither optimism nor pessimism. It’s inventory. Find the physical constraint of your business. Know the lead time. Know who you have in front of you in line, and when you would have had to start to be in operation in 2029. If you can’t answer those four questions, you have no strategy. There is the burn rate.

Andreessen concluded by saying he knew where he was putting his money. These are the stakes, in fact, that are written to be graded, not admired.

1. The binding constraint on incremental frontier-scale capacity in the United States through 31 December 2027 continues to be power availability and grid interconnection, not accelerator supply. Disproved if a published quarter cites delivered accelerator supply, instead of energized capacity, as the reason for a significant buildout delay. Probability .75.

2. High-bandwidth memory, or advanced packaging, margins hold or expand. At least 1 merchant power or GPU-rental category has a visible margin compression cycle from overbuild by 31 December 2028. This tests the second half of the hard lesson and is the claim that is most likely to make me look foolish. If both compress together, then falsified. Probability 0.65

Both enter the register today and will be graded in public whichever way they go.

The hard lesson is not that hardware is valuable. That’s why it’s almost done being an “edge”. Everybody’s figured that out. The hard lesson is that the physical world is the one that sets the clock, it has never once sped up to accommodate a roadmap, it does not care if you have taken the lesson yet.

One disclosure, kept in view: I build and invest in some of the markets I write about, and I say so wherever it is relevant.

Mind Meets Matter. Where intelligence meets its physical limits.