The Latency Stack: Why the AI Economy Runs at Four Speeds

I. Capability: the fast layer, and why it is slower than it looks

Thesis: on the metrics labs report, capability is compounding faster than ever. On the metrics that matter to an economy, it is compounding more slowly, and the headline curves are approaching the limits of the instruments that draw them.

The headline curve

MGI cites the best-known capability statistic in the field. The length of task that frontier AI can complete has doubled roughly every seven months since 2019, and roughly every four months since 2023. The underlying series comes from METR, and it has kept steepening.

• March 2025. METR’s original paper found a doubling time of about seven months over 2019 to 2024 (arXiv).
• January 2026. Its updated methodology put the post-2023 doubling at about 131 days, a little over four months.
• May 2026. For frontier models released since January 2024, the doubling had shortened to about 105 days, roughly three and a half months (OfficeChai). That is about a tenfold increase a year.
• Frontier result. Claude Mythos Preview, evaluated in March 2026, has a 50% time horizon of at least 16 hours, with a wide confidence interval of about 8.5 to 55 hours. Its 80% horizon is about three hours.

The inputs are compounding too. Epoch AI’s trend data (Epoch) shows:

• frontier training compute growing about five times a year;
• the cost of frontier training runs growing about 3.5 times a year;
• the installed stock of AI chips growing about 3.4 times a year;
• compute needed to reach a given level of performance halving roughly every eight months, thanks to algorithmic progress (MIT CSAIL).

And the loop is starting to close on itself. The September 2026 paper co-authored by Hinton, Bengio, Pachocki, Clark and others reports Anthropic data showing AI systems completing about 26% of internal AI R&D work with only high-level supervision by August 2026, against 1% five months earlier (GovAI). If that holds, the capability layer is becoming the first layer in economic history that partly builds its own next version.

On this evidence, capability’s latency is measured in months, and shrinking.

Three reasons the fast layer is slower than it looks

I want to complicate that picture, because the speed of the capability layer is the premise of every investment thesis built on it.

First, the instruments are at their ceilings. METR itself says that measurements above roughly 16 hours are unreliable with its current task suite. The most-cited capability curve in AI is running out of ruler. Rapid gains on a benchmark that is nearly saturated tell you more about the benchmark than about the economy.

Second, economically real tasks move more slowly. SemiAnalysis’s new GDPx benchmark, published on 2 November, uses 206 tasks across 11 domains, built with 14 specialist partners, each with an eight-hour cap (XenoSpectrum summary). The best model, Claude Opus 5.5, scored 12%. Gemini 4 Carbon scored 10.9% and GPT-6 Astra 9.6%. SemiAnalysis is clear that GDPx does not measure how much of the economy AI can automate. But it is the clearest signal yet that the gap between benchmark capability and deployable capability is wide.

The human evidence points the same way. METR’s 2025 randomised trial found experienced developers working 19% slower with AI tools, while believing they were 20% faster (Simon Willison). Its February 2026 update suggests developers are probably faster now, but METR called its new data very weak evidence because of selection effects.

Third, the frontier is getting more expensive to run, not cheaper. The famous price collapse is real at fixed capability. Epoch estimates that the cost of reaching a given performance level falls between nine and 900 times a year, depending on the milestone (Epoch). MIT FutureTech puts the decline at five to ten times a year. But the same MIT analysis finds that the cost of running the most advanced models is rising three to 18 times a year (arXiv). Yesterday’s frontier becomes cheap; today’s frontier stays expensive.

Effective capability latency

Put those three points together and you get a more useful concept than headline capability. I call it effective capability latency: the time it takes for a gain in frontier capability to become a capability that a regulated enterprise can actually deploy, at a cost it can bear, on a task that matters to its P&L.

On that definition, the capability layer is not running in months. It is running in quarters to years, gated by three things.

• Reliability. Can the system hit the 95% or 99% threshold that a bank, a hospital or an airline needs, not merely the 50% at which time horizons are measured?
• Cost at the frontier. Is the newest capability affordable at production volume, or only in a demo?
• Deployment policy. SemiAnalysis attributes part of the GDPx gap to safety classifiers refusing tasks in finance and law. Some of the capability exists but is not released for those uses.

This matters for the rest of the stack. If capability’s effective latency is quarters rather than months, the mismatch with the slower layers is less extreme than the headline curve implies. The real danger then lies elsewhere: in the AI R&D loop, where capability gains compound inside the labs long before they reach the market. That is precisely where governance is weakest, as I argued in The Day Governance Became a Balance Sheet.

The capability layer therefore has two speeds of its own. Internally, inside the R&D loop, it may be the fastest system humans have built. Externally, as experienced by the economy, it is much slower. Most of the public debate conflates the two.

II. Infrastructure: the layer you can pour but cannot rush

Thesis: the binding constraint on AI is no longer chips. It is the grid, the equipment that connects to it and the consent of the people who live beside it. Each of these runs on a multi-year cycle that no amount of capital can compress very far.

The scale of the build

MGI’s paper repeats McKinsey’s 2025 estimate that cumulative global data-centre investment will reach about $7 trillion between 2025 and 2030 (McKinsey, April 2025). The split is telling: about $3.1T for chips and hardware, about $1.3T for power, cooling and electrical equipment, about $0.8T for construction and about $1.5T for non-AI data centres. More than 40% of the total is expected to land in the United States.

The 2026 numbers show the build is running at that pace. On the latest guidance I could find, the four largest hyperscalers plan roughly $720B to $740B of capital spending this year, my own approximate sum of company guidance.

Company	2026 capex guidance
Amazon	about $220B
Alphabet	$195B to $205B, with 2027 signalled above $300B
Microsoft	about $170B to $175B (definitions vary)
Meta	$135B to $145B

Power: the median wait is now five years

The International Energy Agency’s 2026 update reports that data centres consumed about 485 TWh in 2025, up 17% in a year, against roughly 3% growth in total electricity demand. AI-focused data centres grew about 50%. The IEA projects all data centres at around 950 TWh by 2030, roughly 3% of global demand (Rigzone). In the United States, Lawrence Berkeley National Laboratory projects data centres at about 11.8% of national electricity by 2030, within a range of 9.5% to 15.3%, up from 4.4% in 2023 (LBNL).

Demand can rise that fast. Supply cannot. LBNL’s Queued Up report (2026 edition) gives the most important latency number in this essay.

• The median time from an interconnection request to commercial operation was 61 months for projects completed in 2025.
• It was 36 months in 2015 and 22 months in 2008.
• About 2,061 GW of capacity sits in queues.
• Only 13% of the capacity requested between 2000 and 2020 was ever built. About 75% was withdrawn.

The grid’s latency has nearly tripled in less than two decades, during the very period when capability latency collapsed.

The equipment queues tell the same story. GE Vernova reported about 122 GW of gas-turbine backlog and slot reservations at the end of Q3 2026, with power orders up 178% organically (SEC 8-K). Earlier this year it put lead times for new heavy-duty turbines at about three years (Power Engineering). A Wood Mackenzie survey found transformer lead times of about 128 weeks, a supply deficit of about 30% and prices up 77% since 2019 (POWER).

One line from GE Vernova’s chief executive deserves attention. He suggested turbines are not really the gating item. If even the turbine maker thinks something else is slower, the bottleneck has moved further down the stack: to permits, wires and politics.

The market is already routing around the grid. MGI expects more than a quarter of campuses of 500 MW or more to generate their own power by 2030, up from about 1% today. The grid is so slow that the largest buyers of electricity are leaving it.

Chips and memory: the layer that cleared first

The chip bottleneck that dominated 2023 is easing, though not gone. In 2023, advanced packaging limited H100 supply, with lead times of up to 11 months. TrendForce now reports TSMC’s CoWoS packaging capacity heading to about 130,000 wafers a month by the end of 2026 and about 260,000 by the end of 2028 (TrendForce).

Memory has become the price signal instead. Korea’s October 2026 semiconductor exports reached $56.4B, up 259% year on year, with memory alone at $50.1B, up 360% (Seoul Economic Daily). Chips are now nearly half of Korea’s entire export book. This is what a bottleneck looks like when it is priced. Fabs and packaging lines can be added in two to three years: slow next to capability, but fast next to the grid. So the chip layer clears first, and its rents migrate to the slower parts of the stack.

Consent: the newest and least predictable latency

The slowest part of the infrastructure layer is not physical. It is political.

• MGI reports at least 75 data-centre projects, worth about $130B, blocked or delayed by community opposition in the first quarter of 2026 alone.
• Today, 3 November, voters in ten US states decide 41 local data-centre ballot measures, according to Ballotpedia (Smart Cities Dive). Seventeen would prohibit data centres outright.
• S&P counts more than 50 projects cancelled in 2026 amid community opposition (S&P Global).

Electricity prices are the fuel. LBNL found US residential prices up about 5% from 2024 to 2025, and the EIA expects roughly another 5% in 2026. LBNL is careful to say price drivers are state-specific. But public opinion has already decided: in a Consumer Reports survey, 38% blamed AI and data centres for rising bills, and 84% backed a separate rate class for large loads (S&P Global).

Consent has a peculiar property. It is the only infrastructure latency that can go negative. A turbine order, once placed, eventually ships. A ballot measure can stop a project that was already permitted. Consent is not a queue. It is a veto.

Asia saw this film first

Singapore lived the consent-and-power constraint before the United States did. It ran a de facto moratorium on new large data centres from 2019 to 2022, then reopened on its own terms (Reed Smith): a 2022 pilot call that awarded about 80 MW to four operators, a Green Data Centre Roadmap promising at least 300 MW with an efficiency target of PUE 1.3 or better, and a data-centre park of up to 700 MW on Jurong Island announced in October 2025. Data centres already use about 7% of the country’s electricity.

Malaysia followed a similar arc at speed. Johor’s boom brought 51 approved projects worth about RM183B, then higher electricity tariffs for large data centres from July 2025, a dedicated water tariff and, in November 2025, a halt to approvals of lower-tier facilities over water concerns (The Star). Japan and Korea face harder grid limits; NTT has said that in inner Tokyo you might wait five to ten years for power.

The lesson Asia offers is not that the infrastructure layer can be sped up. It is that it can be governed. Singapore turned an open-ended queue into a deliberate, priced allocation tied to efficiency. That is the only lever that genuinely shortens infrastructure latency for the projects that matter: deciding what not to build.

III. Organisations: the layer that rewires slowest

Thesis: almost every firm has adopted AI, and almost none has been rebuilt around it. The organisational layer is not slow because people resist change. It is slow because value requires redesigning work, and redesign is limited by the scarcest resource in any firm: the capacity of humans to review, decide and take responsibility.

Adoption is nearly universal; value is not

• Use. McKinsey’s 2026 State of AI survey finds nearly nine in ten organisations using AI in at least one function (McKinsey).
• Scale. 44% say they are scaling AI across the enterprise, up from 38%.
• Value. Only 37% report any EBIT impact at all, essentially unchanged from 39% a year earlier.
• Leaders. The share of high performers, firms attributing 5% or more of EBIT to AI, is stuck at about 6%.
• Redesign. Only about a quarter of organisations have redesigned workflows around AI. Among high performers, about three-quarters have.

The official statistics are more sober still. The US Census Bureau found 19.8% of US firms using AI in May 2026 (Census), with sector rates that did not change significantly over six months. But a Federal Reserve note found about 41% of workers using generative AI for work, and about 78% of the labour force employed at firms that have adopted AI (Federal Reserve).

That gap is the organisational layer in a single statistic. People adopt at the speed of software. Firms adopt at the speed of org charts.

The disappointments are well documented. MIT NANDA’s 2025 report found that about 95% of organisations saw no measurable return on their generative AI spending (Virtualization Review). Its methodology has been criticised, but its central finding matches McKinsey’s flat EBIT numbers. Gartner predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027.

The productivity J-curve, and the dynamo

None of this should surprise anyone who knows the history of general-purpose technologies. Paul David’s 1990 paper, The Dynamo and the Computer (PDF), showed that electric motors supplied under 5% of mechanical drive in US factories in 1899, and only slightly more than half by the early 1920s. The productivity payoff arrived about four decades after the first central power station, once factories were redesigned around a motor for each machine. MGI makes the same comparison for whole economies: steam took about a century to reach its peak contribution to growth, electricity about 40 years, computers and the internet about 25.

Brynjolfsson, Rock and Syverson formalised this as the productivity J-curve (AEJ: Macroeconomics, 2021). Firms first invest heavily in intangibles: process redesign, training, data and new management practices. Those investments depress measured productivity before they lift it.

Are we past the bottom of the J? The evidence is mixed. US nonfarm productivity rose 2.2% in the year to Q2 2026 on preliminary data (BLS). Brynjolfsson estimates 2025 productivity growth at about 2.7% against a 1.4% decade average and argues the economy is entering a harvest phase (Fortune). Apollo’s Torsten Slok counters that AI is everywhere except in the incoming macro data. No statistical agency yet attributes measured productivity growth to AI.

The real bottleneck: review capacity

Why is the organisational layer so slow? The usual answer is change resistance, skills or culture. Those matter. But this week supplied an unusually clear example of a deeper, structural constraint.

Goldman Sachs has restricted Claude Code and GitHub Copilot to engineers and quantitative strategists. Bankers and traders, who had used them for months, lost access (The Decoder). The firm cited productivity and risk management. In July, its CIO had publicly praised front-office “vibe coding”.

The tools worked well enough that non-engineers used them for months. What did not scale was review. In a regulated firm, every model, script and spreadsheet that touches a client or a position needs someone accountable to check it. Generative AI made producing code nearly free. It did nothing to increase the supply of people qualified to sign it off.

This is the organisational version of Amdahl’s law. If generating work speeds up tenfold but reviewing it does not speed up at all, throughput is capped by review. The organisational bottleneck is not adoption. It is accountability bandwidth. Firms that redesign review, through tiered sign-off, automated testing, agent-level audit trails and clear ownership, will see value. Firms that simply add more generation will drown in unreviewed output.

The Frozen Workforce

The labour market shows the same lag in a more painful form. Brynjolfsson, Chandar and Chen’s “Canaries in the Coal Mine” research found that workers aged 22 to 25 in AI-exposed occupations had seen a 13% relative employment decline by mid-2025 (Stanford Digital Economy Lab). A February 2026 update put the gap at about 16%, with no sign of reversal, and a secondary summary of an August update puts it near 19%, driven by less hiring rather than more exits.

Employers’ intentions run ahead of their actions. In McKinsey’s 2026 survey, 39% expect AI-related headcount declines next year, but only 14% report an actual AI-driven decline in the past year. This is what I have called the Frozen Workforce. Firms are not yet replacing existing staff at scale. They are simply not hiring the next cohort. The cost lands first on the young, who never get the apprenticeship years in which judgement is formed. In a decade, that becomes a shortage of exactly the experienced reviewers the review bottleneck demands.

What fast organisations do differently

1. They redesign workflows, rather than adding tools to old ones. About three-quarters of high performers have done so, against a quarter of everyone else.
2. They invest in review infrastructure as seriously as in generation. Testing, evaluation, audit trails and named owners for every agent.
3. They treat governance as an accelerator. MGI finds only about 30% of organisations at level 3 or higher on a four-level AI governance maturity scale, and about one in three with no formal security check before deployment. Mature governance is what lets a firm deploy at scale without a Goldman-style reversal.
4. They measure HX, not just CX. In my framing, HX = CX + EX: the human experience is the customer experience plus the employee experience. Firms that improve one at the cost of the other find the gains do not stick.

IV. Governance: a control plane slower than the system it controls

Thesis: governance is not merely the fourth layer of the stack. It is the control plane, the only layer that can deliberately speed up or slow down the other three. Today it runs at a higher latency than the systems it is meant to control. In any engineered system, that is the definition of instability.

The control plane problem

Network engineers distinguish between the data plane, which moves traffic, and the control plane, which decides where traffic goes. A basic rule of system design is that the control plane must react faster than conditions change in the data plane. Otherwise, by the time a routing decision arrives, the network it was meant for no longer exists.

The AI economy has inverted that rule. Its data plane, the capability layer, improves in months. Its control plane, governance, legislates in years. MGI lists governance gaps among its key risks, including AI ownership scattered across IT, legal, data, HR, risk and strategy. But it treats governance as one input among many. I think that undersells it. Governance is the one layer whose latency determines whether the speed mismatches elsewhere are survivable.

How slow is the control plane?

The EU. The AI Act was first proposed in April 2021. In July 2026 the Digital Omnibus, Regulation (EU) 2026/1744, delayed its core obligations again (National Law Review). Stand-alone high-risk systems now face obligations from 2 December 2027, and high-risk AI embedded in regulated products waits until 2 August 2028. Obligations for general-purpose AI models were not delayed. That is more than six and a half years between proposal and the main high-risk obligations, during which the time horizon of frontier models rose from minutes to many hours.

The United States. There is no federal AI statute, despite more than 150 bills (TechTimes). Executive Order 14365 of December 2025 created a Justice Department task force to challenge state AI laws (Federal Register). Meanwhile the states keep legislating: NYU’s Center on Technology Policy counted 109 AI laws in 29 states by 1 July 2026 (Tech Policy Press), and another tracker counted 115 in 30 states by mid-October. The result is a control plane that is slow at the centre and fragmented at the edges.

Regulator capacity. The US Center for AI Standards and Innovation had about $10M of funding in fiscal 2026, flat for three years, and a senior lawmaker has said it needs closer to $100M a year (FedScoop). The UK AI Security Institute reportedly has more than three times its staff and about six times its funding. The EU AI Office has over 125 staff and is hiring about 40 more. The frontier labs each employ thousands of researchers. The regulators who oversee them employ hundreds.

History. Slow governance is the norm. Powered flight began in 1903; the US Air Commerce Act came in 1926, about 23 years later (FAA). After the 2008 financial crisis, the US margin rule for uncleared swaps was not published until January 2016, about seven and a half years later (CFTC). The difference now is that the technology being governed changes on a cycle of months, not decades.

Hard and slow, or fast and soft

Asia shows that the governance layer does not have one speed. It has a trade-off.

Jurisdiction	Instrument	Speed	Bite
Singapore	IMDA Model AI Governance Framework for Agentic AI, January 2026	Fast; a living document	Voluntary
Singapore	MAS Guidelines on AI Risk Management, issued 7 October 2026	Consultation to issue in about 11 months	Supervisory; core sections apply from October 2027
Korea	AI Basic Act, in force January 2026	Moderate	Binding, but maximum fine of about US$21,000, after a grace year
Japan	AI Promotion Act, 2025	Fast	No monetary penalties
China	AI content labelling measures, in force September 2025	Fast	Mandatory national standard
EU	AI Act, high-risk obligations	Slow; delayed to 2027 and 2028	Binding, with heavy penalties

The pattern is stark. The faster the instrument, the softer it tends to be. Fast governance is usually unenforceable; enforceable governance is usually slow. The interesting work is in instruments that are both: supervisory guidelines that regulated firms must follow, procurement conditions, listing rules, contractual clauses and insurance requirements. None needs a new statute. All can move at something closer to the speed of the systems they govern.

Where the control plane is weakest

Most of the governance above regulates deployment: what firms do with AI. Almost none of it reaches the AI R&D loop inside the frontier labs, the internal fast lane described in Section I. That is the subject of The Day Governance Became a Balance Sheet. Anthropic’s amended S-1 told investors it may soon cross its own threshold for AI-accelerated AI development, and that it has historically noticed such acceleration only afterwards. The fastest part of the capability layer is governed by the slowest instruments in the stack.

In the language of this essay: the control plane has the highest latency precisely where the data plane is fastest.

V. Where the layers grind: bottlenecks, arbitrage and Asia

Thesis: a desynchronised stack produces three things at once: rents for whoever owns the slowest stage, stranded capacity for whoever overbuilt the fastest, and risk that accumulates in the gap. The winners of the next five years will not be the fastest movers. They will be the ones who cut latency where it is highest.

Amdahl’s law for the AI economy

Gene Amdahl’s 1967 insight was simple. If only part of a task can be accelerated, the overall speedup is capped by the part that cannot. Written out: overall speedup = 1 ÷ ((1 − p) + p ÷ s), where p is the share of the work that gets faster and s is how much faster it gets. As s grows without limit, the speedup approaches 1 ÷ (1 − p). If AI can accelerate 60% of a value chain infinitely, the whole chain still gets at most 2.5 times faster, because the other 40% did not move.

Applied to the Latency Stack, the implication is uncomfortable for anyone whose strategy is simply “more capability”. Each further gain in model capability is worth less than the last until the grid, the organisation and the rulebook catch up. Capability without the slower layers is a sports car in a traffic jam.

Effect one: rents migrate down the stack

1. 2023 to 2024: accelerators and packaging. GPU supply was gated by advanced packaging. The chip designers captured extraordinary margins.
2. 2025 to 2026: memory. HBM and DRAM became the price signal. Korea’s memory exports in October were up 360% year on year.
3. 2026 onward: power equipment, grid and land. GE Vernova’s power orders rose 178% organically in Q3. Developers who already hold grid connections and powered land own the scarcest asset in the stack.
4. Next: consent and review. Communities that can say no, and the qualified people who can sign off AI-generated work, are slow to expand and impossible to manufacture.

The recent market moves fit this pattern. Since 6 October roughly $3.5T to $3.7T has come off chip and memory stocks (Yahoo Finance), while orders for turbines and grid equipment hit records. I would not read that as the end of AI demand. I would read it as the market repricing which layer is now the bottleneck.

Effect two: stranded capacity and Ghost GDP

When a fast layer outruns a slow one, capacity strands. MGI gives the cleanest example: Microsoft reported holding AI chips it could not deploy because it lacked powered data-centre space. MGI also lists overinvestment and stranded capacity among its key risks, citing the railway and internet booms. This is the mechanism behind what I have called Ghost GDP: measured activity, in the form of capex, construction and imports, for a capability whose output has not yet arrived. MGI notes US imports of AI-related goods are on track for about $750B in 2026. Ghost GDP is not fake. But it is fragile, because its value depends on the slower layers eventually catching up.

Effect three: risk pools in the gaps

• Between capability and organisations: unreviewed output, shadow AI and the kind of front-office tooling Goldman has now withdrawn.
• Between capability and governance: the AI R&D loop, accelerating inside labs governed by self-imposed policies.
• Between infrastructure and consent: projects built ahead of local acceptance, now stranded by moratoriums and ballot measures.

MGI names the scenario that connects them: a major AI cyberattack or governance failure that triggers a backlash and slows adoption across the board. In Amdahl terms, a shock in the slowest layer can reduce the throughput of the whole system.

Latency arbitrage: where the value goes

Slow stage	Latency today	What cuts it	Who captures the value
Grid connection	About 61 months median in the US	On-site generation, priority queues for flexible loads, pre-permitted sites	Power developers; landowners with grid access
Power equipment	About 2.5 to 3 years	Standardised designs; capacity expansion	Equipment makers; long-dated buyers
Local consent	Unbounded, and can reverse	Community benefit agreements; separate rate classes; water-neutral designs	Operators who earn a licence to operate
Organisational review	Months per workflow	Evaluation, testing and audit tooling; tiered sign-off; named agent owners	Firms that redesign work; governance and assurance providers
Regulation	Years	Supervisory guidelines, procurement conditions, assurance standards	Jurisdictions that are fast and credible

This is what I mean by Long-AND, not Short-OR. Being long capability alone is a bet on the fastest layer, whose marginal value is falling. The durable position is long capability and long the slow layers that unlock it.

What this means for Asia

Korea is leveraged to the bottleneck that clears fastest. Memory is the current price signal, and chips are nearly half of Korea’s exports. When packaging and memory capacity catch up, Korea’s trade balance will feel the rotation first. The strategic question for Seoul is how to move up the stack, into power, grid equipment and AI deployment, before the memory cycle turns.

Singapore’s advantage is low governance latency with credibility. It ran one of the first serious data-centre moratoriums, reopened with priced and efficiency-linked allocation, and issued supervisory AI guidelines in about 11 months. In a stack where the control plane is the slowest layer, a jurisdiction that can govern quickly and credibly is scarce. I think Singapore can be the control plane of the region: the place where assurance standards, model evaluation and agentic governance are set, tested and exported.

Malaysia shows that consent constraints are global. Johor’s turn from boom to water-driven approval limits took less than two years. Japan’s binding constraint is the grid, which makes its AI trajectory a question of energy policy more than technology policy.

China plays a different game. MGI notes that China leads in open-weight models and industrial deployment, while the US narrowly leads in frontier models and high-end compute. Those are bets on the organisational and deployment layers, not the frontier. In Latency Stack terms, that may prove the smarter allocation.

Asia’s public is more willing. MGI reports that 55% of people in China and 54% in India expect AI to improve the job market within three to five years, against 18% in the United States. If consent is becoming the slowest infrastructure latency in the West, Asia’s comparative optimism is an under-appreciated economic asset. It is also a responsibility, because optimism spent carelessly is how backlash begins.

VI. The panel: how my usual voices would push back

A note on method. These are not new interviews. Each entry sets out a figure’s position from the public record, quoting only where I could source the words, followed by the strongest critique that position makes of this essay. Where I reconstruct a likely response, I say so.

The MGI authors (Shubham Singhal, Arvind Govindarajan, Tanguy Catlin, María Jesús Ramírez, Xavier Rigby). Their paper argues that leaders must learn to manage at multiple speeds. Their likely critique: the Latency Stack over-weights governance; for most firms, organisations are the binding constraint, and governance is one input that shapes them. My reply: for an individual firm, that is right. For the system, governance is the only layer that can set the speed of the others, which is why I treat it as the control plane rather than an input.

Erik Brynjolfsson, Stanford. He co-authored the productivity J-curve and argues the economy is entering a harvest phase. His likely critique: the essay is too pessimistic about the organisational layer; the J-curve is turning. My reply: I hope he is right. But his own Canaries research shows the cost of the transition landing on young workers, and a harvest that freezes the next generation’s apprenticeship will leave the review bottleneck worse in a decade.

Daron Acemoglu, MIT. His widely cited projection is that AI will add only about 0.5% to productivity over a decade (The Next Web). His likely critique: the slow layers may never fully catch up, because many tasks are hard to automate reliably. My reply: the GDPx results support part of his view. That is precisely why I distinguish headline capability from effective capability latency.

Torsten Slok, Apollo. His line that AI is everywhere except in the incoming macroeconomic data is the sharpest summary of the Ghost GDP problem. His likely critique: the capex boom is itself the risk; if returns do not show up, correlated investment across chips, power and data centres will unwind together. My reply: agreed. The recent chip drawdown is a preview.

Fatih Birol, IEA. His position is that there is no AI without energy, and his agency’s 2026 data shows data-centre demand growing about six times faster than total electricity demand (Eco-Business). His likely critique: power may be the master constraint for the next decade. My reply: for the infrastructure layer, yes. But even with abundant power, an organisation that cannot review AI output and a regulator with ten million dollars will still cap the system.

Scott Strazik, GE Vernova. He has said his company sold many 2030 turbine slots, while suggesting turbines are not really the gating item. The implicit challenge: much of the infrastructure latency is administrative and political, not industrial. That strengthens this essay’s argument that infrastructure latency is a governance problem in disguise.

Dario Amodei, Anthropic. His September essay argues the industry must slow the pace of capability improvement, because AI is increasingly building the next generation of AI. His likely critique: the essay calls the capability layer slower than it looks, when the internal R&D loop is accelerating faster than anyone outside can see. My reply: we agree more than it seems. Externally the layer is slower than the hype; internally it may be faster than our instruments. The governance gap sits between the two.

Gary Marcus. He has long argued that today’s systems are unreliable in ways benchmarks hide. His likely critique: the 12% GDPx score shows the capability layer is far slower than even my effective-latency framing allows. My reply: he may be right about reliability. But a capability that is 12% today on a benchmark that did not exist last month is exactly the kind of fast-moving baseline the slower layers cannot plan around.

VII. Coming next in this series

1. Part II: Powered Land. The infrastructure layer in depth: grid queues, turbines, transformers, self-generation, consent, and the Singapore and Johor playbooks.
2. Part III: Accountability Bandwidth. The organisational layer: review capacity, the Frozen Workforce, HX = CX + EX and how high performers redesign work.
3. Part IV: The Control Plane. The governance layer, building on The Day Governance Became a Balance Sheet: why the slowest layer governs the fastest loop, and what fast-but-enforceable governance looks like.

VIII. Notes on sources

MGI’s figures are taken from its September 2026 discussion paper as published on its website. Several figures rest on secondary reporting and are flagged in the text, including METR’s May 2026 doubling estimate, the GDPx scores, Anthropic’s AI R&D share and Japan’s power-wait estimate. The hyperscaler capex total is my own sum of company guidance. State AI law counts vary by tracker. Full source links appear inline throughout.

Luke Soon is an AI futurist and ethicist, and author of Genesis: Human Experience in the Age of Artificial Intelligence and Synthesis: The Superintelligence Protocol.