Bain’s latest technology report puts a number on the AI buildout that should make every operations leader pause: the industry needs $6 trillion in annual revenue by 2031 to justify what it’s spending on data centres. Annual AI infrastructure spending could hit $1.5 trillion by 2031. Meta’s Ohio facility alone is projected to cost $200 billion by 2030. Someone pays for that, and it won’t be the hyperscalers absorbing it quietly.

That someone is you, through vendor pricing, licence tiers, and how much model capacity you can actually afford three years from now. Bain’s David Crawford calls absorption speed, the pace at which companies put AI to work, the new competitive variable. Below, what those numbers mean for your roadmap, your budget, and the cost of waiting.

A $6 Trillion Bill Is Being Written Against Your AI Roadmap

Bain’s 29 September report breaks the $6 trillion down, and the breakdown is the interesting part. Enterprise productivity, the thing every vendor pitch is built on, accounts for $1 trillion to $1.4 trillion of it. New product development, search, advertising, autonomy and physical AI, carries $4.2 trillion. Productivity gains cover roughly a fifth of what the buildout needs.

“The debate today is fixated on employee productivity. The economics of AI infrastructure demand trillions in new revenue beyond productivity gains,” said David Crawford.

For a plant or quality function, that gap sets your pricing conditions. Vendors chasing a revenue target that productivity alone cannot hit will push you toward higher tiers, metered inference, and premium model access. Slow buyers get the worst terms, because they negotiate last and from zero internal evidence.

What Bain Actually Counted: $4.2T in New Products vs $1.4T in Productivity

Read the split as a forecast of where vendor attention goes. Four fifths of the required revenue comes from things that do not exist yet as mature markets. One fifth comes from making work you already do cheaper. That asymmetry decides which products get engineering headcount and which get maintenance mode.

Why physical AI and autonomy sit in the biggest revenue bucket

Search, advertising, autonomy and physical AI share one property: they create new transactions rather than shaving cost off existing ones. A cheaper QA review saves you money. An autonomous inspection system that nobody was buying in 2024 is net new spend. Only the second category scales fast enough to cover data centres whose size and cost double every 12 to 16 months.

For manufacturing, that is good news in one narrow sense. Physical AI means vision systems, robotics and machine-level autonomy get serious investment because the industry needs them to pay the bill. Expect capability in those areas to improve faster than in generic back-office copilots.

What the industry needs is a wave of innovation that will dwarf what mobile and cloud unlocked.

What a $1.4T productivity ceiling implies about enterprise AI pricing

Bain caps enterprise productivity revenue, spanning software development, sales, marketing, customer service and IT operations, at $1 trillion to $1.4 trillion. That is the entire addressable pool for the category most of your current pilots sit in. A ceiling that firm means vendors cannot grow that line by volume alone. They grow it by price, by seat count, and by moving capability into higher tiers.

The counterweight is absorption speed, which Bain calls the new competitive variable, with leading AI labs putting upwards of $9.75 billion into engineering models that help companies assimilate faster. Vendors are spending real money to shorten your deployment cycle because slow customers do not generate recurring revenue. Buyers who can absorb quickly will be courted. Everyone else gets the list price.

Doubling Every 12 to 16 Months: The Cost Curve Behind the Headline

Bain’s figure for data centres is not driven by inflation or land prices. Facility sizes and costs are doubling roughly every 12 to 16 months. That is a compounding curve on the supply side, and it runs underneath every licence renewal and per-token price you will negotiate between now and 2031.

Look at what the annual spend actually buys. It is not one line item. It covers new facilities, GPU upgrades, memory and networking equipment, and three of those four are replacement cycles rather than one-off construction. Upgrade cycles do not end. They re-bill.

This matters because most multi-year AI business cases quietly assume the opposite. The working assumption in budget decks is that inference gets cheaper every year, so a pilot that is marginal at today’s prices becomes comfortable by year three. Per-token list prices on older models have fallen, which makes the assumption feel safe. It is not.

Cost per unit of capability is the number that moves, and it moves unevenly. Older models get cheaper. The frontier models you actually need for complex inspection reasoning, multi-document deviation review, or anything touching regulated documentation stay expensive, because they sit on the newest and most capital-intensive hardware. The cheap tier and the capable tier are drifting apart, not converging.

For buyers, three consequences follow. First, assume flat-to-rising cost for frontier access and falling cost only for commodity tasks, then design your architecture around that split. Second, treat model choice as a cost engineering decision, not a procurement footnote: route the routine 80 percent of volume to cheap models and reserve the expensive tier for cases that genuinely need it. Third, stop writing three-year business cases that depend on a price drop you cannot contract for.

If your AI roadmap only clears its hurdle rate because inference is projected to be 70 percent cheaper in 2029, you do not have a business case. You have a bet on someone else’s capital expenditure schedule, and that schedule is doubling.

Absorption Speed Is the Competitive Variable Bain Named, and Most Factories Fail It

Bain defines absorption speed as the pace at which companies can actually put AI to work, and calls it the “new competitive variable”. Leading AI labs are putting upwards of $9.75 billion into engineering models that help companies assimilate faster. That is a lot of money spent solving a problem that is not technical on your side of the fence.

Model capability stopped being the constraint a while ago. What stops a vision inspection model or a deviation triage assistant from going live is almost always your data, your documentation, and your change control process.

The four things that actually slow absorption in manufacturing

• Data that lives in people: Tribal knowledge about why a line drifts on Tuesdays is not in any system, so the model learns a partial process.
• Undocumented workflows: You cannot automate a decision path nobody has written down. Mapping it is the real project.
• Validation ambiguity: In regulated environments, no one owns the question of what qualifying an AI-assisted inspection step requires. The pilot stalls in limbo.
• No production owner: Pilots are run by innovation teams. Production workflows need a line manager whose numbers depend on it.

Fix those four and your absorption speed roughly triples without touching the model. Ignore them and you will buy better models every year and ship nothing.

Measuring your own absorption speed in 90-day cycles

Use one metric: elapsed days from use case identified to running in production with a named owner. Not piloted. Running, with someone accountable for the output. Most manufacturers who measure this honestly land somewhere north of nine months.

A 90-day cycle in a quality function looks like this. Days 1 to 20, pick one inspection or deviation workflow and document it end to end. Days 21 to 50, assemble and clean the data, agree the validation standard with QA before building anything. Days 51 to 80, build and run in parallel with the human process. Days 81 to 90, cut over or kill it. Then start the next one.

How to Build an AI Programme That Survives the Capex Correction

Nobody outside Bain knows whether the $6 trillion arrives. Two scenarios exist: the revenue shows up and pricing stays roughly rational, or it doesn’t and the correction gets passed down through licence tiers, rate limits, and deprecated model endpoints. A programme built properly survives both.

Three rules do most of the work. Keep your workflows portable, so swapping a model provider is a configuration change and not a rebuild. Tie every deployment to one operational number you already report to the board: scrap rate, inspection hours per shift, deviation closure time. And fund use cases that pay back inside one budget cycle, not ones justified on a 2031 horizon you cannot forecast.

Three AI investments that hold their value regardless of vendor pricing

• Clean, labelled operational data: Your defect images, deviation records, and process logs are yours. They get more valuable as models get cheaper, and no vendor repricing touches them.
• Internal absorption capacity: The people who can scope a use case, validate output, and run a change control cycle. Bain called absorption speed the “new competitive variable” for a reason, and leading labs are spending upwards of $9.75 billion engineering around it.
• Abstracted integration layers: One interface between your MES, QMS, and whatever model sits behind it. Build this once and model churn becomes a procurement decision, not an engineering project.

What loses value fast: deep single-vendor customisation, proprietary agent frameworks you cannot export, and pilots scoped around a model’s current context window. Those are bets on today’s pricing holding, and nothing in the cost curve suggests it will.

Our position is simple. The firms that compound advantage between now and 2031 are the ones absorbing fast and cheaply right now, building the data, people, and plumbing that survive any pricing scenario. Waiting for the market to settle means arriving late, paying more, and starting the absorption work your competitors finished three years earlier.

Source: thenationalnews.com