The AI race moved from capability to commercialization
In short
- The claim: capability is no longer the bottleneck — commercialization is. Companies buy AI at scale and have difficulties finding the return in gross margin, headcount or cycle time. The scarce skill is getting a model into a business and keeping it there, and that is commercial work.
- The deadline: roughly $3 trillion has gone into AI infrastructure since ChatGPT launched in Nov 2022, and Cahn's arithmetic puts the end-customer revenue needed to justify it at $1.5 trillion a year. Coding is the one place the money has arrived; everywhere else, adoption has run ahead of return.
- Why pilots stall: not the model. Five gates — someone owns the number and closes the old road; a narrow, measurable workflow; a decision about what humans still decide; a named check before shipping; an agreed expansion path.
- The regional part: Southeast Asia has the widest version of the problem — the highest piloting band anywhere and almost no non-adopters, yet six in ten report under 5% profit effect. Little legacy technology makes buying cheap; owner-run firms mean nobody below the owner changes a process that works.
- What I do about it: read the problem from the commercial side, then build. da-system.ai is six working tools in a live P&L; the persona system cut a manager task from days to 2.5 hours.
The deadline
Capability was the story for years, and rightly: if the model could not do the thing, nothing else mattered. It is not the constraint now. Generative AI reached 53% adoption in three years, faster than the PC or the internet, counting open-weight tools alongside the frontier APIs. US consumer surplus hit an estimated $172 billion a year by early 2026, up from $112 billion, with most of those tools still free.
Estimated annual US consumer surplus from generative AI
$ billion per year
Figure 1. Estimated annual US consumer surplus from generative AI. Up 54% in a year; median value per user roughly tripled. Source: Stanford AI Index 2026, Economy chapter.
Since 2023 David Cahn of Sequoia has tracked what the ecosystem must earn to justify what it builds: Nvidia's projected fourth-quarter data centre run rate, doubled for non-chip cost, doubled again for margin across the cloud provider and the AI company. September 2023, $200 billion. June 2024, $600 billion. July 2026, $1.5 trillion, roughly $3 trillion cumulatively since ChatGPT launched. Open weights do not escape this: they run on the same fleet, and self-hosting moves the bill rather than removing it.
Annual end-customer revenue required to justify AI capex
$ billion per year, revenue required against hyperscaler data centre capex
Figure 2. Revenue required against the capex driving it. Cahn's second derivation doubles 2026 hyperscaler data centre capex of about $764 billion and lands on the same $1.5 trillion. He expects the chip-based version to understate the requirement over time, since it misses TPU and ASIC spend and rising memory and construction cost. Revenue-required figures: Sequoia and Cahn. Capex series: Goldman Sachs Global Investment Research, combined hyperscaler data centre capex ($156B 2022, $254B 2024, $443B 2025, $764B 2026E). Source: David Cahn, July 2026.
Cahn points at the one place the money has arrived. AI coding has a far clearer path to revenue than three years ago, and Anthropic has gone from a standing start to annualized revenue in the tens of billions, with quarterly profit above $1 billion. Then commercialization has not failed. It worked completely in the one market where the buyer needed no persuading. Everywhere else sits an organization full of people who did not ask for this.
The evidence, and what is wrong with it
The number everyone quotes comes from MIT's Project NANDA: roughly 95% of enterprise AI pilots show zero measurable P&L effect. I used it and stopped. Futuriom could not reproduce it from the report's own exhibits and asked for the data or a retraction. The report then offers NANDA, which runs a paid corporate membership, as the way across the gap it describes.
There is a better number, from April this year. Stanford's 2026 AI Index has 64% of organizations reporting improved innovation. On profitability, 36% improved against 36% no effect; on revenue growth, 33% against 39%.
AI impact on organizational measures over the past year, 2025
% of respondents
Figure 3. Self-reported effect of AI, 2025. Respondents already using AI in at least one function. "Worsened" never exceeds 2% on any measure, so this is stalling, not damage. Source: Stanford AI Index 2026, fig 4.3.5, charting McKinsey survey data, 2025. Three of nine measures shown.
Companies are getting better at inventing things and no better at earning from them. S&P Global agrees: the share of companies abandoning most of their AI work rose from 17% to 42% in a year. The figures are contested. The pattern is not.
The five gates
Pilots rarely fail because the model underperformed. Gartner's April 2026 Hype Cycle puts agentic AI at the Peak of Inflated Expectations, with only 17% of organizations actually running agents, and cost, security and governance controls maturing fastest. Its standing forecast of over 40% of agentic projects cancelled by end-2027 names escalating cost, unclear business value and inadequate risk controls. Only one of the three touches the technology, and on the chart below technical limits rank second at 38%, well behind security and risk at 62%. Five gates, roughly in the order they surface.
In McKinsey's Southeast Asia survey of 330 companies, lack of executive sponsorship ranks last of nine barriers at 3%; globally it also ranks last, at 9%. Yet 55% of high performers say senior leaders take real ownership, against 29% of everyone else. BCG puts 70% of AI value in people and process, and tells leaders to back three or four priorities, not a hundred use cases.
Main obstacles to reaching fully scaled agentic AI, 2025
% of respondents, global
Figure 4. Barriers to scaling agentic AI, global. Chart source: Stanford AI Index 2026, fig 3.3.10, charting McKinsey survey data. "Other" (2%) and "None" (1%) omitted. The regional barrier ranking and the high-performer split quoted above come from McKinsey, EDB and Tech in Asia. Exhibit 3 gives the regional ranking, where lack of executive sponsorship is last at 3%. The high-performer exhibit defines high performers as companies above US$250M revenue attributing more than 5% of EBIT to AI (n=29 against 184, a small base).
The barrier ranked last is what separates the winners. That is distance, not hypocrisy: budget was approved; words were said at the offsite; nobody checked whether it was still in use six months later. Permission changes nothing while the old way works. The sponsor's job is to close the old way.
If you cannot say what the task costs today, you cannot show it costs less tomorrow. Anthropic's guidance on building agents starts in the same place: check that the task is worth the run, that its error cost is tolerable, and that the model can do it at all before building a loop around it. OpenAI's field guide adds the sequencing: set evals to establish a baseline first, then choose the model against it. Broad scope with no baseline makes a pilot impossible to disprove and impossible to fund.
This is the one I got wrong. We ran a trial producing short-form video for public figures with large followings and no clear point of view. Each video was fine alone; in sequence they fell apart, because nothing carried across. We had automated the making and never settled who owned the judgement. The fix: build the persona first, make the content plan an output of it, and keep a person in between. BCG says the same at company scale: rethink workflows end to end rather than bolting a model on.
Not review in general: a specific test, run before output leaves the team, with a defined response when it fails. That is an eval, and Anthropic's engineering team describes the alternative precisely: flying blind, debugging reactively, unable to separate a real regression from noise. OpenAI treats guardrails as layered rather than singular, with a defined human handover when the system cannot finish. Enterprises rank the same thing first among barriers: 62% name security and risk. Most often skipped, most often fatal: trust breaks on the first bad output and does not come back.
Success criteria set at signing, pricing that lets the pilot convert, and a rollout where each stage earns the next: yourself, the team, the company, customers. Claude Code was built in that order: narrow evals first, harder behaviours later, monitoring and A/B tests added as it scaled. OpenAI says the same: start small, validate with real users, add capability only once the earlier stage holds. Escalating cost, Gartner's first cause, is what happens when a pilot scales before anyone agreed what scaling meant.
The regional part
Most of the good writing comes out of the US, for sound reasons: that is where the models and the largest budgets sit. Evidence from markets where adoption ran ahead of redesign is thinner on the ground, and that is the part I can add.
Southeast Asia is the region where everyone has started and almost nobody has finished. It has the widest piloting band anywhere at 35%, and not a single firm reporting no use, against 6% globally and 14% across the rest of Asia Pacific. Yet six in ten report under 5% effect on operating profit despite putting 11% to 40% of technology budget into AI.
Adoption against reported profit effect
% of respondents
Adoption stage by region
Effect on EBIT, Southeast Asia (n=330)
Figure 5. Adoption against money. Bases: global n=2,084, US n=701, Southeast Asia n=330, Asia Pacific ex China and India n=187. Regional figures are composite, weighted by enterprise size and GDP share across six economies. Source: McKinsey, EDB and Tech in Asia, February 2026, exhibits 2 and 6. Two separate exhibits, placed side by side here.
So the gap here is not smaller because adoption is faster. It is wider. Two facts from the same report, usually read apart, explain it: less legacy technology, so layering something new on top is cheap, and more family-owned businesses, where direction comes from the owner. Buying is fast, and nobody below the owner can authorize changing a process that works. Adoption without redesign: gate three failing because gate one never closed, since nobody below the owner can close the old road.
The consumer side rhymes. Work by Cube with impact.com and dentsu across 2,400 shoppers found 24% use AI to discover products and 28% to research them, but influencers drove 67% of purchases and family or friends score highest on influence. AI is where people look, not where they decide. Deciding runs through creators and affiliates, worth about $70 billion, or 32% of regional e-commerce in 2026. So the commercial lever here is not AI-generated content competing with creators; it is AI put behind creators, amplifying the manual judgement that already converts. That is the gap the persona system was built for, and it is where consumer AI actually gets paid for.
Agentic AI adoption across business functions, Southeast Asia
% of respondents scaling or fully scaled, n=330
Figure 6. Agentic adoption by function, Southeast Asia, % scaling or fully scaled. The functions closest to revenue sit at the bottom. Global figures run far lower: scaling plus fully scaled reaches roughly 14% in IT and single digits nearly everywhere else, and 68% of firms report no agent use at all in marketing and sales. Regional numbers are self-reported and the report says executives may be overstating maturity. Nearly nine in ten regional companies plan to experiment with agents during 2026. Sources: McKinsey, EDB and Tech in Asia, exhibit 5; Stanford AI Index 2026, figs 4.3.7 and 4.3.8.
What I am doing about it
Find the problem where it sits, read it from the commercial side rather than the model side, then build. da-system.ai is six working tools on the Anthropic API, built inside a P&L I own. The persona system shows the clearest before and after: from two or three days of a manager's time down to about two and a half hours, across ten-plus managers and a hundred-plus creators, with that cohort's followership up 17% in three months. No control group on the last figure, so treat it as directional.
Before that I built commercial teams across nine markets and started the key account enterprise business from nothing in six of them.
If you are building the commercial side of an AI company in Asia Pacific, or you think part of this argument is wrong, email me or find me on LinkedIn. I would like to hear it.