United States – The artificial intelligence industry will need to generate $6 trillion in annual revenue by 2031 just to justify the scale of compute investment now underway, according to Bain & Company’s latest report. Much of that sum, the firm argues, will have to come from entirely new sources of value, not merely from productivity gains inside existing businesses.
Existing AI applications will continue to expand, according to the report. Consumer AI, monetised through subscriptions and advertising, and enterprise AI, spanning software development, sales, marketing, customer service and IT operations, could together bring in between $1.2 trillion and $1.8 trillion.
That still leaves a $4.2 trillion gap, which Bain expects new categories of innovation to fill. Four stand out: model providers displacing search engines and folding in advertising revenue; autonomous vehicles, drones and industrial automation opening up new product lines; physical AI, from simulations to digital twins and robotics, reshaping research and manufacturing; and entirely new markets, such as drug discovery, mental health and energy generation, built on what Bain calls “abundant intelligence”.
“The debate today is fixated on employee productivity. The economics of AI infrastructure demand trillions in new revenue beyond productivity gains,” said David Crawford, Chairman of Bain’s Global Technology Practice.
“What the industry needs is a wave of innovation that will dwarf what mobile and cloud unlocked. AI infrastructure is being built well ahead of the demand curve and funding it sustainably will require adding approximately 1% to the annual global GDP growth rate,” Crawford continued.
Hardware strikes back
Surging demand for AI compute has reversed years of underperformance in hardware. Semiconductor and hardware stocks climbed at a 24% compound annual rate between 2020 and 2026, more than four times the 6% posted by software firms over the same period. High-bandwidth memory, advanced packaging and custom silicon, particularly application-specific integrated circuits, are driving much of that growth, the report found.
The co-development of memory and logic silicon has made it harder for buyers to switch suppliers, while heavy investment in high-bandwidth memory capacity has left double data rate memory and NAND comparatively starved of capital, a trend that could tighten supply and push up prices for smartphones and PCs.
Custom chips, once a niche pursuit, are also gaining ground. Hyperscalers and AI-native companies are increasingly designing silicon tuned to their own workloads, a shift Bain links to the arrival of training, inference and agentic workloads large enough to justify bespoke chip design.
Meanwhile, natural disasters, geopolitical tension and export controls are pushing companies to diversify their supply chains, with foundries responding in kind, particularly in logic chips.
“This is a dynamic time for players in the hardware sector. Supply chain and procurement strategies are increasingly sources of competitive advantage as investing in supplier capacity, long-term agreements, equity investments in the supply chain, and multi-vendor, multi-geography sourcing are now very high on the C-suite agenda,” said Anne Hoecker, Global Head of Bain’s Technology practice.
“Also, product strategy choices are changing. The old model where one vendor innovates and sells to everyone else is changing, and a wave of verticalization and semi-custom design are becoming more prevalent,” Hoecker continued.
Cybersecurity faces a watershed moment
Cybersecurity has climbed to the top of the CISO agenda, driven partly by high-profile incidents involving frontier AI model testing. Bain calculates that AI has cut the time needed to mount a typical cyberattack from roughly four weeks to about 18 hours, a shift compounded by the spread of AI agents.
Poor “agentic housekeeping” is a glaring vulnerability, and the absence of an off-the-shelf fix is leaving many companies hesitant rather than proactive in choosing a build, buy or partner strategy.
Supply chain exposure is another growing worry, according to the survey, which found that companies struggle to track how vendors deploy AI across the life of a contract, including midcycle changes and fourth-party risk. Traditional oversight, built around infrequent questionnaires, is proving too slow for an environment where vendors update models and software daily.
In response, leading firms are overhauling how they remediate vulnerabilities, lifting remediation budgets by double-digit percentages in many cases and reassigning between 20% and 25% of cybersecurity staff to handle alerts generated by AI-powered scans. DevOps teams are being drawn into the effort too, while the most advanced organisations are concentrating resources on the hardest risks to fix, including those tied to legacy platforms, network layers and SaaS vendors.
Absorption speed becomes the new edge
According to the report, how quickly a company can put AI to work or “absorption speed” is emerging as a fresh competitive battleground. Leading AI labs are pouring more than $9.75 billion into forward-deployed engineering models designed to help enterprise customers integrate AI faster, while vendors race to build the application and infrastructure layers that turn raw model intelligence into business results.
Even as competition intensifies and token prices fall, Bain pushes back on the idea that large language models are becoming commodities.
“We see the more likely scenario as a continuum of frontier and mature models. New and unproven cases will initially favor frontier models, then likely migrate to lower-cost alternatives as the cases mature
Frontier providers will push toward more fundamental challenges, capturing value where superior intelligence matters most. Meanwhile, lower cost, specialized, and optimized models will efficiently serve more common tasks, providing the services and support that enterprise customers need,” Crawford said.
“The industry end state is far from settled. We are likely headed into a segmentation between frontier models and trailing models or an expanding definition of models rather than a classic commoditization pattern,” Crawford continued.
Choosing the right model for the right task is only part of the challenge, the report suggests; building the engineering management systems to make that choice effective is another. A separate survey of 293 senior technology leaders found expectations running high, with respondents anticipating a 148% improvement in release-cycle speed and a 95% rise in developer productivity within one to two years, far outpacing the 20% to 27% gains most are currently seeing across key metrics.
While AI accelerates coding itself, it pushes bottlenecks downstream into review, coordination, quality and governance, meaning gains made at the task level can simply relocate friction rather than eliminate it unless the whole software development lifecycle is redesigned.
The companies pulling ahead, the research concludes, tend to do three things: structure the knowledge their AI agents rely on, build quality and trust directly into their workflows, and treat the software development lifecycle as a product to be continuously improved, rather than a fixed process.
