AI Must Chase $6 Tn in Annual Revenue by 2031 to Fund Its Build-Out

The artificial intelligence industry would need to earn about $6 Tn a year by 2031 to sustain its projected spending on chips, data centres and power, Bain & Company said in its annual technology report on September 29. A year ago, the Boston-based consultancy projected $2 Tn a year by 2030.

The target has risen because spending has. Bain now estimates annual investment in AI infrastructure could reach $1.5 Tn by 2031, covering everything from new data centres and computing capacity to hardware upgrades for installed GPUs, memory and networking equipment. This forecast assumes spending will eventually settle at about a quarter of industry revenue, a ratio Bain describes as ambitious but reasonable given major cloud providers’ historical track record.

The five biggest spenders are Microsoft, Google’s parent company Alphabet, Amazon, Meta, and Oracle. According to the report, their combined capital expenditure could reach $780 Bn this year, nearly five times their spending levels just three years ago.

Existing applications will not cover this investment. Bain estimates that by 2031, consumer subscriptions and advertising could bring in $200–400 Bn annually, while enterprise adoption could yield another $1–1.4 Tn in gains for AI providers. Even at the upper end of those estimates, that leaves a $4.2 Tn deficit that must come from markets that are only beginning to develop.

The consultancy identifies four distinct sources to bridge this gap. Model providers could earn $100–200 Bn by integrating advertising into AI search, while autonomous vehicles and industrial automation could unlock about $400 Bn. Physical AI, which spans digital twins, simulation, and robotics, could represent a $900 Bn opportunity, assuming a 10% reduction in research, development and manufacturing costs across selected industries.

However, how much of that value will directly translate into revenue for AI providers remains uncertain. The remaining revenue will need to emerge from entirely new applications in fields such as AI-driven drug discovery, mental health support and new ways of energy generation.

Bain’s argument extends much beyond making existing work cheaper. The firm envisions AI rendering treatments for rare diseases commercially viable, accelerating the discovery of advanced materials for batteries and semiconductors, and driving breakthroughs in complex scientific research fields like fusion energy. While these applications represent massive pathways for new economic activity, their ultimate commercial scale remains unproven.

“What the industry needs is a wave of innovation that will dwarf what mobile and cloud unlocked,” said David Crawford, chairman of Bain’s global technology, media and telecommunications practice. He noted that AI infrastructure is currently being constructed well ahead of the actual demand curve. Funding this massive build-out sustainably, Crawford added, would require adding around 1% to global GDP’s annual growth rate.

Table 1
A Breakdown of AI’s $6 Tn Revenue Requirement by 2031

Revenue, Spending & Value Pool	Bain Estimate
Annual AI Infrastructure Spending	$1.5 Tn
Annual AI Market Revenue Needed to Sustain It	~$6 Tn
Consumer AI Subscriptions & Advertising	$200-400 Bn
Enterprise AI Gains to Providers	$1-1.4 Tn
New Revenue Still Needed, from the Following:	~$4.2 Tn
Search & Advertising	$100-200 Bn
Autonomous Vehicles & Industrial Automation	~$400 Bn
Physical AI	~$900 Bn
New Products (e.g. drug discovery, energy generation)	Not quantified

Source: Bain & Company, Technology Report 2026 (September 29, 2026)
Note: The $6 Tn assumes capital spending of about 25% of industry revenue. The $4.2 Tn gap assumes consumer and enterprise AI generate the upper estimate of $1.8 Tn. The physical AI opportunity assumes a 10% reduction in R&D and manufacturing costs across selected industries. It is not a direct forecast of provider revenue. The four categories do not constitute a fully quantified allocation of the gap. Bain does not put a figure on new products. 

Most Companies Are Still Learning to Use AI

While some businesses are already recording concrete financial gains, true value remains concentrated. Bain & Company reports a 10% to 25% growth in earnings before interest, taxes, depreciation, and amortisation (EBITDA) among pioneering companies that have undertaken broader AI transformations. Yet the firm’s findings suggest that as many as 90% of enterprises remain bogged down in deploying isolated tools and pursuing narrow applications, resulting in limited impact on top-line revenue or bottom-line earnings.

Achieving meaningful scale extends far beyond merely purchasing technology. Organisations partnering with Bain have discovered that every dollar allocated to software and infrastructure demands an additional four dollars spent on restructuring people and internal processes. Redesigning operational workflows, modernising enterprise data systems and shifting employee responsibilities represent substantial, capital-intensive investments in their own right.

To bridge this operational gap, AI providers are aggressively deploying capital to help customers navigate the transition. Bain tracks more than $9.75 Bn in committed capital across collaborative initiatives involving Anthropic, OpenAI, Microsoft, Amazon Web Services (AWS) and Google Cloud to accelerate enterprise deployment. A core component of these commitments includes sending specialised engineering teams to work directly on-site with customers, ensuring AI is deeply integrated into their day-to-day operations.

The primary constraint shifting across the tech landscape is no longer the technology itself, but how rapidly organisations can absorb new capabilities. Citing a 2025 Massachusetts Institute of Technology study, Bain & Company identifies weak executive sponsorship, arduous organisational change, subpar user experiences and cultural resistance to new tooling as the primary obstacles. Merely purchasing access to a more capable model leaves these systemic, operational bottlenecks entirely unresolved.

A massive monetisation opportunity lies in the cross-system work employees do across legacy software suites—reconciling disparate data, chasing internal approvals and manually co-ordinating decisions involving billing, sales and customer support. Bain estimates a highly addressable $100 Bn US market for agentic software that can execute this interstitial work. With software vendors currently capturing a mere $4-6 Bn of that pool, technology companies have an unprecedented opening to charge directly for completed jobs that enterprise customers currently pay their human employees to co-ordinate.

Bain projects a similarly scaled opportunity across Canada, Europe, Australia and New Zealand, expanding the combined addressable market to roughly $200 Bn. However, the geographic and functional scope for automation varies substantially. While Bain estimates that 40% to 60% of tasks within customer support and R&D workflows could be seamlessly automated, that figure drops to 20% to 30% for specialised legal work, where operational errors carry severe consequences.

Ultimately, friction points like access to usable clean data, the technical complexity of connecting legacy systems and the structural need for physical intervention will dictate how much workload these agents can reliably assume.

How this multi-billion-dollar revenue pie will be distributed among market participants remains fluid. Bain expects frontier, top-tier foundation models to retain their value for highly complex, novel applications, while routine corporate tasks are likely to migrate towards cheaper, specialised alternatives.

A substantial portion of industry profits may also shift towards the software layers that connect frontier models to proprietary corporate data and internal business processes, turning integration architectures into the next major battleground for tech competition. These orchestration systems manage the underlying information, toolsets and enterprise permissions required for autonomous agents to act, while dynamically routing workflows between competing models.

Bain also expects the competitive boundary to keep changing, as capabilities currently managed by surrounding software layers are increasingly absorbed into core foundation models.

The rise of AI agents creates another dilemma for software vendors: how much of their data and functionality should be accessible to external networks.

The emergence of the Model Context Protocol (MCP) has provided a standardised framework for agents to plug directly into enterprise business software. While opening application programming interfaces (APIs) can embed a legacy vendor deeper into customer workflows, it also lets corporate clients execute complex tasks entirely outside the vendor’s native user interface, threatening to cannibalise high-margin seat licences. Conversely, restricting access risks rendering a software platform obsolete and driving clients towards more accessible, open-ecosystem competitors.

Bain argues that software vendors must resolve these structural exposure strategies before designing their monetisation models.

Software Investors Face a Different Calculation

For software investors, the economics are becoming harder to assess. Bain & Company notes that median annual revenue growth among US-listed software companies has slowed from around 20% to roughly half that rate. AI also introduces substantial computing costs each time a product is used, while charging by usage or results can make revenue less predictable than conventional subscriptions. Private equity buyers need evidence that new AI sales can cover their delivery costs and sustain margins.

Median net revenue retention among US-listed enterprise SaaS companies, which tracks revenue from existing customers after expansion and losses, has fallen by about nine percentage points since 2021. However, software assets still changing hands can command high prices, partly because owners are bringing their strongest businesses to market.

Meanwhile, post-pandemic technology buyouts, including investments made in 2020-22, have so far generated returns below pre-pandemic averages, although many have yet to be fully realised.

Bain says investors need to examine how AI could alter or replace customer workflows a business serves, alongside the cost of adapting its products. Companies also need to distinguish revenue from established machine-learning products, AI assistants and autonomous workflow products. Each carries different costs, margins and competitive risks, which a single recurring-revenue figure can obscure.

AI Savings Put Service Providers under Pressure

Technology services companies face a related squeeze. Customers are demanding efficiency improvements up to 20% at contract renewals and awarding new work on the assumption that delivery costs will fall by at least 30% and sometimes by as much as 50%, Bain says. With contracts often running for three to five years, providers can find themselves committing to productivity gains they have yet to achieve.

Bain estimates potential productivity improvements of 15-45% across different service lines, but says few providers have built repeatable commercial models for delivering AI at scale. After all, charging for staff time becomes harder to reconcile with a technology intended to reduce the labour required. Therefore, pricing based on results and sharing the savings with customers are gaining ground.

Established providers retain an advantage in their deep knowledge of clients’ systems, data and business rules. However, turning that knowledge into new revenue requires more than adding AI tools to existing services. Bain sees clear opportunities for firms that specialise in particular industries, rebuild their delivery around AI and develop the engineering and commercial teams needed to turn proposals into funded projects.

The Hardware Bill Is Already Being Paid

For hardware suppliers, the spending boom is already translating into tangible revenue, with much of the manufacturing in Asia. Bain found that the market capitalisation of hardware and semiconductor companies expanded at a CAGR of 24% between 2020 and 2026, sharply outperforming software’s 6% growth.

The report highlights high-bandwidth memory (HBM)—the stacked chips that feed data to AI processors—as one of the fastest-growing industry segments, alongside advanced packaging and custom silicon. HBM revenue surged from about $4 Bn in 2023 to an estimated $77 Bn in 2026, while South Korea’s SK Hynix and US-based Micron posted record gross margins of 75–85% in each of the past two quarters.

HBM’s position within the tech supply chain is undergoing a structural shift. Because stacked memory is co-developed alongside a foundation logic layer, switching suppliers is much harder for the chipmakers that buy it. This technical integration gives Samsung, SK Hynix and Micron an active role in chip design and production, fundamentally strengthening their leverage with buyers.

However, the infrastructure boom carries a distinct cost for end consumers. Because the three dominant memory manufacturers have diverted most of their new capacity to high-margin HBM, investment in conventional DRAM and NAND chips has stalled, worsening component shortages and driving up retail prices for smartphones and PCs, Bain noted.

AI demand is also reshaping the processor design landscape. Large cloud providers and frontier AI firms are increasingly commissioning bespoke chips tailored to their specific workloads, creating major opportunities for specialist designers and their manufacturing partners. Bain identifies these application-specific integrated circuits, or ASICs, as the fastest-growing category of data-centre computing silicon.

Massive, recurring workloads across model training, inference and agentic applications can now easily justify the steep upfront capital expenditure required for custom designs. Buyers also face powerful incentives to curtail escalating compute costs and power consumption. Consequently, Bain projects that custom silicon volumes could soon eclipse Nvidia’s GPU shipments, although Nvidia is poised to retain the lion’s share of total market revenue.

Advanced packaging, the phase that integrates discrete chips into functioning modules, is also capturing a growing share of industry value. Gross profit from semiconductor packaging climbed to $21.5 Bn in 2025 from $12 Bn in 2022, with Taiwan Semiconductor Manufacturing Company (TSMC), the world’s largest contract chipmaker, and HBM vendors capturing the bulk of this new value, according to the report.

But this extreme geographic concentration remains the industry’s primary structural vulnerability. Bain characterises Taiwan as both the sector’s largest supply source and its greatest bottleneck, prompting chip buyers to aggressively secure alternative fab locations and secondary vendors. Natural disasters, geopolitical friction and export restrictions continue to threaten both component availability and pricing structures.

Manufacturers are responding differently to these risks. While logic foundries plan to build about 49% of their new capacity for 2026-30 outside their home countries, memory makers are retaining all but 8% of theirs domestically, Bain’s data shows.

“Supply chain and procurement strategies are increasingly sources of competitive advantage,” said Anne Hoecker, global head of Bain’s technology practice.

The report also highlights long-term purchasing agreements, direct investment in supplier capacity and equity stakes across the supply chain as vital ways for buyers to lock in access.

These shifting dynamics are fundamentally altering product development workflows. More enterprises now want to work directly with chipmakers on bespoke designs tailored to their precise requirements. Therefore, Bain sees a broader structural shift towards closer integration and customised products as buyers demand greater control over the technologies their businesses depend on.

Governments are also adding to the pressure. Now that they treat leading-edge silicon as vital strategic infrastructure, Bain expects them to push harder for domestic or allied production through subsidies, export controls and local-content rules. Escalating East-West tensions are also forcing multinationals to re-evaluate their sourcing and sales corridors, prompting many to ask whether they need entirely separate strategies for Chinese and Western markets.

TSMC has committed $265 Bn to expanding semiconductor manufacturing and advanced packaging in Arizona, according to the company and the US Commerce Department. Bain adds that recent conflicts in the Gulf have introduced a novel stress test for infrastructure planning. Fabrication facilities and data centres must now be designed around exposure to military disruption, maritime choke points and regional security, in addition to conventional access to capital and physical resources.

Singapore and Malaysia Target More Market Share

Singapore is already capitalising on the strategic migration of logic foundries beyond their domestic markets. Classed within the “rest of world” category in Bain’s capacity datasets, the city-state commands around 10% of global semiconductor output and 20% of semiconductor manufacturing equipment production, according to the Economic Development Board.

The regional footprint expanded significantly on September 28, when VisionPower Semiconductor Manufacturing Company (VSMC), a joint venture between Taiwan’s Vanguard International Semiconductor and Dutch chipmaker NXP Semiconductors, commenced risk production at its $6.7 Bn fabrication facility in Tampines. The site will manufacture speciality silicon and the vital silicon interposers required for advanced packaging systems.

Securing memory manufacturing investments will prove a steeper challenge. Bain’s projection that memory makers will retain all but 8% of planned wafer capacity domestically indicates that new fabs will remain heavily concentrated in home markets. Nevertheless, Singapore retains a crucial foothold in this segment. Micron is currently constructing a NAND wafer fab scheduled to begin production in the second half of 2028, and an HBM advanced-packaging facility is slated to expand capacity from the first half of 2027.

Meanwhile, Malaysia’s competitive edge remains anchored in the supply chain’s backend. The country processes approximately 13% of global chip assembly, testing and packaging, according to the Malaysian Investment Development Authority (MIDA). However, multinational corporations execute the vast majority of this volume. Industry executives note that domestic firms hold virtually no market share in advanced packaging, the precise high-margin segment Bain identifies as capturing the industry’s new value.

“Malaysia is at zero,” said Tan Eng Tong, representing the Malaysia Advanced Packaging Consortium, during an industry conference in May.

Kuala Lumpur is moving aggressively to alter this. The government has pledged at least RM25 Bn in fiscal support to operationalise its National Semiconductor Strategy, targeting a 7% global market share in advanced packaging by 2035. As part of this push, the country announced RM92 Mn in research grants in May for the consortium’s five domestic firms. These companies are currently developing an advanced packaging prototype that integrates next-generation HBM4 memory with a test chip. This project includes a dedicated pilot line for final assembly and testing.

Table 2
Where Singapore and Malaysia Sit in the AI Hardware Build-Out

Market	Position and Planned Expansion	Key Data Point	Pressure Point
Singapore	Wafer fabrication, semiconductor equipment; HBM packaging capacity expansion planned from H1 2027	~10% of global semiconductor output; ~20% of global semiconductor equipment production	Land, labour, power
Malaysia	Assembly, testing, packaging; advanced packaging push	~13% of global assembly, testing and packaging	Fragmented local capabilities across the advanced packaging chain
Johor (Malaysia)	Hyperscale and colocation data centres	1,110 MW live; 8,542 MW pipeline	Power, water, approvals

Sources: Singapore Economic Development Board (EDB); Malaysian Investment Development Authority (MIDA); Knight Frank, Data Centre Atlas 2026; Micron; Malaysia Advanced Packaging Consortium, as reported by The Edge Malaysia, EDB Singapore

Note: Pipeline capacity includes proposed developments and is not operational capacity.

Power, Water and Land Set the Limits

Hosting the build-out brings its own strain. Bain projects $5-6.5 Tn of cumulative data centre spending by 2030, adding nearly 150 GW of computing capacity and nearly tripling global data centre capacity in just five years. Most of the growth will come in the US, although capacity in every region will rise by double digits. However, power, hardware and skilled-labour shortages, along with slower public approvals, are holding back the expansion.

Those constraints can reinforce one another. Bain says adding grid capacity takes four years or more, while building a 1 GW data centre typically takes one to four years even on a site with existing power. Chips, memory and networking equipment are being booked years ahead, leaving operators without long-term supply agreements at a disadvantage. Decisions taken now will determine how much capacity is available several years from now.

The scale of the investment could also force consolidation, leaving fewer companies able to finance the largest projects. Bain sees scope for shared power procurement and jointly funded grid upgrades, while governments could move beyond subsidies into more direct roles as financiers, co-ordinators or owners.

Meeting electricity demand will require several approaches. Renewable generation, storage and onsite power can provide nearer-term capacity. Major new nuclear projects and orbital data centres remain longer-term possibilities, with timelines and economics that offer little certainty for the immediate build-out. Community acceptance will also matter, as concerns over water use, electricity prices and noise increasingly affect approvals.

Johor shows the pressure. The Malaysian state has 1,110 MW of live data centre capacity, a pipeline of 8,542 MW and a vacancy rate of 0.7% in shared, or colocation, facilities, according to Knight Frank’s Data Centre Atlas 2026. New projects must now explain their power source, water supply and cooling to win approval, and the state has banned two categories of highly water-intensive facilities.

Singapore, which paused new data centre approvals from 2019 to 2022, now releases capacity in small, conditional rounds. In August, it provisionally allocated 200 MW on Jurong Island to Digital Realty, Equinix, Keppel Data Centres and ST Telemedia Global Data Centres, under a call requiring at least half of applicants’ power to come from green sources.

Cybersecurity Adds to the Cost of Scaling

Security adds another cost. Bain estimates that AI has compressed the time required for a typical cyberattack from about four weeks to 18 hours. The report points to incidents involving tests of frontier models, the most advanced AI systems, underscoring that defences built around human decision-making need to adapt to attacks operating at machine speed.

Defensive AI tools can uncover weaknesses faster, but companies then have to fix them. Bain reports vulnerability alerts increasing by as much as eightfold. Leading companies are raising remediation budgets by double-digit percentages, while some are redirecting 20-25% of their cybersecurity staff to address findings from AI-powered scans.

Development teams are also being pushed to spend more time finding and fixing vulnerabilities.

The proliferation of AI agents creates further exposure. Bain says many businesses would struggle to establish how many agents they have, who owns them and what they are permitted to do. Controls to stop an agent’s activity are another concern. Waiting for a single vendor to supply a complete solution can leave those weaknesses unaddressed.

Risks also enter through suppliers and their own technology providers. Companies need to track changes in how vendors use AI throughout a contract, including deployments introduced after the original agreement. With models and software changing daily, periodic questionnaires provide an increasingly incomplete picture. Bain calls for continuous monitoring and stronger contractual controls over new AI deployments.

Legacy platforms, network systems and external software providers are among the most concentrated and difficult exposures identified in the report. Customers cannot fix some independently. Beyond clearing the immediate backlog of vulnerabilities, businesses face a longer programme of identifying agents and AI workloads, strengthening threat detection and establishing clear responsibility for their use.

Turning AI Gains Into Revenue

The revenue side depends on how fast companies put AI to work. Bain’s survey of 293 senior technology leaders found median gains of 27% in release-cycle speed and 20% in developer productivity compared with pre-AI performance.

Respondents expect far bigger gains within one to two years. They see release-cycle speed improving by 67% (148% above today’s level) and developer productivity by 39% (up 95% from now). The wider challenge is to turn improvements across business functions into the $1-1.4 Tn in annual revenue for AI providers that Bain estimates enterprise AI adoption could generate by 2031. All of it is part of a broader $6 Tn total market requirement needed to fund the global infrastructure build-out.

But faster execution of individual tasks does not automatically produce faster delivery. The survey found that developers completed around 21% more tasks, while review time increased by about 91%. Companies can generate more code and still struggle to deliver finished products if approvals, quality checks and cross-team co-ordination cannot keep pace.

Bain identifies three structural changes needed to overcome those operational bottlenecks. First, organisations must document engineering knowledge so that autonomous agents can consume it. Second, they need to build automated quality checks directly into development. Third, they should continuously improve the full software production pipeline.

Also, human supervision remains central. Only 6% of respondents envisioned fully autonomous development, while 41% expected oversight to vary according to the risk profile of the work.

The tech workforce is shifting rapidly alongside the technology. A separate Bain survey of 155 respondents found that the share of engineering organisations with traditional pyramid structures fell from 66% to 29% over two years. Developers now expect to spend a third of their time directing AI agents within the next two years. Unsurprisingly, AI fluency has emerged as the most sought-after engineering capability, cited by 62% of respondents.

That squeezes the next generation of engineers. Junior engineers made up about a third of teams and now account for less than a fifth. AI is taking over the routine work through which they used to learn, just as companies prize senior judgement more highly.

Decision-making has yet to catch up. Bain says 70% of companies still make decisions centrally, so that approvals move no faster even as code is produced more quickly.

For Singapore and Malaysia, the $6 Tn macro question is a matter of value-chain positioning. Both nations are already deeply embedded in the foundational hardware supply chain, maintaining critical footprints in wafer fabrication, semiconductor equipment and conventional packaging.

The near-term test is whether they can move into higher-margin work such as advanced packaging and high-bandwidth memory. That will depend on securing enough power, water and land for fabs and data centres, while the revenues needed to sustain the build-out remain uncertain.

The longer-term question is whether their own companies can earn part of the $4.2 Tn in new revenue that Bain says must come from AI applications, from physical AI to agentic software, and not just host the hardware that powers them.