Is AI Profitable? The $1.4 Trillion Scorecard Nobody Wants to Talk About

Despite massive investment, AI profitability remains unevenly distributed and largely unproven at scale. Most enterprise AI spending generates productivity gains too diffuse to show clearly on balance sheets, while infrastructure costs remain enormous. A handful of platform providers capture concentrated returns, leaving the broader $1.4 trillion investment landscape yielding modest, often unmeasurable financial returns for most organizations deploying it.

Technology Strategy · AI Economics · May 2026
Andrew Baker · Group CIO, Capitec Bank

For the last three years, the technology industry has spoken about artificial intelligence with almost religious certainty. CEOs describe it as inevitable, venture capital firms describe it as transformational, and analysts describe it as the next industrial revolution. Every quarterly earnings call contains some variation of the phrase “AI first strategy,” delivered with the same tone companies once used for “mobile first,” “digital transformation,” or “the metaverse.” Entire corporate roadmaps are being rewritten around the assumption that AI will fundamentally reshape productivity, economics, and competition over the next decade.

Meanwhile, something strange is happening underneath the noise. The entire industry is spending astonishing amounts of money, yet very few people are asking the most uncomfortable question in technology right now: is AI actually profitable? Not “valuable,” not “important,” not “strategic,” and not “world changing.” Profitable.

That distinction matters because the modern AI conversation has become extraordinarily good at measuring activity while becoming surprisingly vague about measuring return. Somewhere between hyperscaler capital expenditure, startup funding, enterprise modernisation programmes, power infrastructure expansion, and semiconductor demand, humanity may have already committed more than $1.4 trillion toward the industrialisation of artificial intelligence. Despite that extraordinary spending wave, the underlying economics remain remarkably unclear.

1. Is AI Actually Profitable?

There is a website called isaiprofitable.com. It exists for one reason: because the honest answer to that question is so uncomfortable that someone decided the question needed its own URL. Built by Michael Tan Sikorski and updated monthly from SEC filings, earnings calls, leaked financials, and estimates from Bloomberg, the Wall Street Journal, The Information, and Epoch AI, the site gives you a single unambiguous scoreboard. The answer is simple and delivered without apology. No. Not yet. Not at scale. Not in the way the press releases describe.

The site’s industry total as of May 2026 is an estimated $1.4 trillion in cumulative AI spend against $718 billion in revenue across all tracked companies. But to read that table clearly you need to understand that it contains two fundamentally different kinds of business sitting next to each other without distinction. One group is digging for gold: the hyperscalers, model builders, and cloud platforms spending enormous capital in the hope that AI applications generate returns. The other group is selling shovels: the companies whose products every gold digger must buy regardless of whether they find anything.

The gold diggers: companies betting on AI applications and infrastructure

Cumulative all-time estimates. Source: isaiprofitable.com, May 2026.

The shovel sellers: companies profiting from every dollar the gold diggers spend

Full year 2025 actuals from company filings.

Market valuation: sorted by P/E ratio, highest to lowest

*P/E ratios as of 23 May 2026. *Private valuations from latest funding rounds.*

Reading that table from top to bottom is reading the market’s verdict on the AI economy in one view. The two highest P/E ratios belong to companies that manufacture physical objects: a Dutch lithography machine maker and an American chip designer. They are being valued above every software and cloud company in the list because the market has concluded that their physical monopolies are more durable than subscription revenue from platforms burning capital to stay competitive. The pure-play model companies at the bottom have no P/E at all because they have no earnings, and the private valuations assigned to them are bets on future winner-take-most outcomes rather than any present financial reality.

This data is entirely consistent with enterprise research. MIT’s NANDA initiative found that despite $30–40 billion of enterprise investment, 95% of organisations studied were seeing zero measurable return. Forbes Research’s 2025 AI Survey found fewer than 1% of executives reporting significant ROI. An IBM global study of 2,000 CEOs put the share of AI initiatives delivering expected ROI at 25%, with 56% of those CEOs reporting zero significant financial benefit. Morgan Stanley surveyed the S&P 500 in Q4 2025 and found only 21% of those companies could cite a measurable AI benefit at all. These are not fringe studies from AI sceptics. They are mainstream institutional research conducted at scale, and they are telling the same story.

What you are looking at is the strangest economic phenomenon of this decade. Capital is accelerating toward an asset class that is, by every available measure, not yet generating returns commensurate with the investment, and the acceleration is increasing rather than slowing.

2. The Problem Nobody Defines Properly

One of the reasons AI debates become so emotionally charged is that people use the word “profitability” to describe completely different things without realising it. Nvidia is profitable from AI demand. Cloud providers are increasingly profitable from AI infrastructure consumption. Consultancies are profitable from AI implementation programmes. Some startups are profitable from AI wrappers, automation tooling, and niche orchestration layers. None of that automatically means AI itself is economically efficient for the organisations adopting it.

These are entirely different layers of the stack, yet the industry constantly merges them into a single success narrative. Infrastructure profitability is not the same thing as enterprise profitability. A cloud provider selling enormous volumes of GPU capacity can be wildly successful while the companies consuming that capacity struggle to generate meaningful margin improvements from the workloads running on top of it.

The distinction becomes even more important when examining real operational environments. A bank using AI to reduce average support call handling time by fifteen percent may generate measurable economic value because the operational workflow is tightly constrained and measurable. A development team using agentic coding tools to double source code generation may appear dramatically more productive on paper, but if review overhead, architectural inconsistency, platform instability, and defect remediation all rise simultaneously, the profitability equation becomes significantly more complicated. The AI industry has become extremely good at presenting activity as evidence of value, but activity and profitability are not the same thing.

3. Seven Hundred Billion Dollars and a Prayer

The five largest hyperscalers have collectively committed between $660 billion and $690 billion to capital expenditure in 2026, with roughly 75% directly tied to AI infrastructure. Goldman Sachs projects that AI related capital expenditure will represent over 3% of US GDP in 2026, up from 1.6% in 2025. Bank of America estimates that the key hyperscalers will spend approximately 90% of their operating cash flow on capex this year, and have responded to the shortfall by issuing debt, with data centre debt issuance doubling to $182 billion in 2025 alone. Amazon is projected to turn free cash flow negative this year. These are historically cash generative businesses burning their balance sheets to fund infrastructure whose return horizon nobody can confidently specify.

The depreciation arithmetic sharpens the picture further. Given that AI assets typically depreciate at around 20% per year, the implied annual depreciation expense facing the hyperscalers exceeds $400 billion, which is more than their combined profits in 2025. The Nvidia gear purchased today will be obsolete before it is fully depreciated, yet the buying continues at an accelerating pace.

“FOMO has proven a stronger incentive than poor stock performance as hyperscalers have prioritised being involved in the AI arms race over their current shareholders.”

Goldman Sachs analyst David Covello, Fortune, May 2026

4. AI Has Reintroduced Industrial Economics Into Software

One of the most important changes happening underneath the AI boom is that software no longer behaves like traditional software. For decades, software had beautiful economic properties because once written it could scale globally at near zero marginal cost. Distribution was effectively free, replication was effortless, and infrastructure requirements remained relatively modest. That economic model helped create the modern internet economy.

AI changes those assumptions dramatically because modern large scale models behave more like industrial infrastructure than traditional software products. Suddenly power consumption matters again. Cooling systems matter again. Fibre routes matter again. Geographic proximity to compute matters again. Water supply matters again. Semiconductor supply chains matter again. Governments are now discussing sovereign AI capacity in the same way countries once discussed energy independence or telecommunications infrastructure.

Traditional Software vs AI Economics

That shift is profound because AI is not simply another software wave layered on top of existing economics. It is the reintroduction of physical infrastructure constraints into computing itself. The economics start looking less like SaaS and more like heavy industry disguised as software.

5. OpenAI’s Maths Are Broken

OpenAI hit a $20 billion annualised run rate by the end of 2025 and crossed $24 billion by April 2026, which is genuinely historic growth at any scale. Beneath that headline, the company posted a $13.5 billion net loss in just the first half of 2025, and internal documents leaked to the Wall Street Journal project a $74 billion operating loss in 2028 alone. Microsoft’s revenue share data implies that OpenAI burns approximately two dollars of cost for every dollar of inference revenue, before research and development, sales, marketing, or overhead. OpenAI started serving advertisements in the free tier in February 2026, which is the playbook every consumer internet company runs when subscriptions cannot carry the cost base.

The entire AI infrastructure supply chain is a closed, recursive financing loop. OpenAI drives Microsoft’s spend. Microsoft drives Nvidia’s orders. Nvidia drives data centre build-outs. If any node in that loop slows, all nodes slow simultaneously. When OpenAI missed its internal revenue and user targets in late April 2026, Oracle fell roughly 3% in a single session and Nvidia, Broadcom, and AMD followed lower. The largest single buyer of AI compute may be growing slower than the infrastructure build-out assumed. That gap between revenue growth and contracted compute spend is the most important risk in the entire AI investment thesis, and it is not being discussed proportionately to its significance.

6. The Subprime AI Crisis

Ed Zitron, writing in March 2026, offered the most structurally precise description of AI’s economic fragility yet published. His framing is worth engaging with directly because it names the mechanism rather than just the symptom.

The parallel he draws to the 2008 subprime mortgage crisis is not about scale but about architecture. Subprime mortgages created the illusion of sustainable demand by hiding the real cost in teaser rates that would later adjust upward. Adjustable rate mortgages let millions of borrowers believe they could afford houses they could not, because the payment in year one bore no relationship to the payment in year three. The housing market inflated on the back of that manufactured demand, and when the rates adjusted and the real cost became visible, the entire structure collapsed. Zitron’s argument is that AI subscriptions are the teaser rate mortgages of the technology industry.

The mechanics are worth tracing. Nvidia sells GPUs to data centre developers and hyperscalers, and this is the only link in the entire chain that is genuinely profitable. Data centre developers raise debt to fund build-outs and rent compute to AI labs, and most of those data centre projects are unprofitable even with a paying customer. CoreWeave, the largest and best funded AI compute provider, reported an operating margin of negative 6% and a net loss margin of negative 29% in 2025. AI labs then rent that compute to train models and serve inference, selling subscriptions to end users at prices that bear no relationship to what it actually costs to serve them. Anthropic was allowing users to burn up to $8 of compute for every dollar of their subscription. OpenAI users were burning over $2,000 of compute on a $200 monthly plan. AI startups building on top of AI lab APIs were in the same position: Cursor, Replit, Lovable, and Augment Code were all burning several dollars of token cost for every dollar of subscription revenue. Every link in the chain was selling at below cost, funded by the expectation that either venture capital would remain permanently available or that some future economic event would make the unit economics work.

What Zitron documents in precise operational detail is what happened when the labs tried to start recovering their costs. Beginning in June 2025, Anthropic and OpenAI introduced priority service tiers for enterprise API customers, requiring upfront commitments and guaranteed throughput minimums. Cursor was forced to abandon per-request pricing entirely. Replit moved to effort-based pricing. Augment Code moved to a credit model that users found so opaque and punishing that one enterprise customer burned $15,000 in tokens on a $250 monthly plan before anyone noticed the bill. In March 2026, Anthropic introduced peak hour rate limits, and users on the $200 monthly Max plan found themselves hitting 60% of their session limits after four prompts. The reactions, documented across Reddit and Twitter in real time, were not the reactions of people who understood they had been using a subsidised product. They were the reactions of people who had built their workflows and professional identities around a service and were watching it change under them without warning, which is exactly what happens when a teaser rate expires.

The structural problem this exposes is deeper than a pricing dispute. AI labs have spent years training their most engaged users to expect near-unlimited access at flat monthly rates, disconnected from the actual token economics of what they were consuming. The most engaged users are also the most expensive users, because model costs increase with the complexity and length of interactions. An agentic coding session where Claude reads a codebase, implements a feature, runs tests, and iterates through failures consumes tokens at a rate that can easily exceed $100 of compute in a single session. The $200 monthly subscription was never designed to cover that. Neither was the $20 one. The subscriptions were designed to attract users, and the economics were deferred to a future round of venture capital. That deference now has a maturity date, and the rate adjustments are the beginning of the reset.

Zitron’s most pointed observation is about the impossibility of the exit. AI startups cannot raise prices to economic cost because users trained on subsidised rates will not accept a 10x increase in what they pay, even if that increase simply reflects what the service actually costs. They cannot reduce rate limits significantly without destroying the product experience that drove growth. They cannot go public easily because their financials are indefensible to institutional investors who read income statements rather than growth narratives. And they cannot be acquired at current valuations by the hyperscalers who are themselves burning cash on the same infrastructure bet. Cursor, valued at $29.3 billion, has turned roughly $3.36 billion of venture capital into approximately $1 billion of revenue. There is no price at which that business makes sense for an acquirer unless the acquirer believes the growth trajectory will change what the unit economics refuse to support.

The subprime analogy is imperfect, as Zitron acknowledges, because the scale is smaller and the contagion mechanisms are different. There are no AI backed CDOs infecting pension funds in the way that mortgage backed securities did. But the underlying psychological architecture is the same: a product was sold at a price that created the appearance of value and accessibility, demand was manufactured on the back of that subsidised pricing, and the infrastructure was built to serve demand that only exists because of the subsidy. When the subsidy ends, the demand will contract, and the infrastructure will be exposed as having been built for a market that was never real at its stated price.

7. The Bull Case, and Why It Deserves a Serious Hearing

Before accepting the sceptic position in full, it is worth steelmanning the argument on the other side, because the people authorising these losses are not all acting on fear alone.

The first part of that thesis is the deliberate subsidy argument. The hyperscalers are not accidentally losing money on AI. They are pricing AI below cost to capture market share and lock in consumption patterns before the category matures, in the same way AWS priced cloud compute below cost for most of its first decade until it owned the category and then expanded margins aggressively. Whether you find this convincing depends on whether you think AI is more like cloud, where one provider can own the category, or more like a commodity input where competition keeps margins structurally thin regardless of early positioning.

The second part is the inference cost curve. Epoch AI research shows that LLM inference prices have been falling at rates approaching 200x per year when accounting for both pricing and efficiency improvements since January 2024. The cost of achieving GPT-4 level performance fell from approximately $60 per million tokens in early 2024 to under $1 by early 2026, a collapse of over 98% in roughly two years. The bull case argument is that falling inference costs open demand curves that do not yet exist.

There is a significant complication, however, and it lives in the shift toward agentic AI. A chatbot sends one message, gets one response, and stops. An agent works differently. It runs a reasoning loop with tool calls, file reads, edits, validations, and checks, and each step in that loop sends the entire accumulated context to the model. By step twenty of an autonomous task, you are paying for the same system prompt and conversation history twenty times over. Engineering teams running agentic AI in production are consistently finding the AI bill becoming the second largest line item on the engineering ledger after salaries within ninety days of enabling agents. Gartner projects that inference on frontier models will cost 90% less by 2030 than it does today, but also predicts that cheaper tokens will not translate to cheaper enterprise AI precisely because agentic models require far more tokens per task. The enterprise average AI budget has already grown from $1.2 million per year in 2024 to $7 million in 2026, even as per-token costs have collapsed.

The third part of the bull case is the unrecognised revenue question. For Amazon, Alphabet, and Microsoft, AI is not primarily a product they sell. It is infrastructure that defends and extends cloud franchises that generate vastly more revenue than any AI line item. Microsoft’s AI business surpassed an annualised revenue run rate of $37 billion, up 123% year on year, as of Q3 FY2026. Azure surpassed $75 billion in annual revenue for the first time, with AI contributing an estimated 13–16 percentage points of that growth. Google Cloud grew 63% in Q1 2026, with backlog nearly doubling quarter on quarter to over $460 billion, and Search revenue grew 19% in the same quarter with Pichai explicitly attributing it to AI experiences driving queries to all time highs.

The honest read is that the scorecard understates both the losses and the returns. It understates the losses because the depreciation on $400 billion of annual capex will compound into operating margins for years. It understates the returns because the revenue that AI is generating inside cloud growth, search volume, and enterprise platform expansion is real but systematically misattributed. What that does not change is the core problem facing every enterprise buyer: you are purchasing from vendors whose pricing is subsidised by capital markets rather than unit economics, and the eventual normalisation of that pricing is a dependency in your cost model that you probably have not explicitly acknowledged.

8. Productivity Does Not Automatically Become Profit

The strongest argument for AI is usually framed around productivity. Developers write more code, analysts generate reports faster, marketers create content at enormous scale, support agents respond more quickly, and legal teams produce drafts in minutes instead of days. The capability leap is real, and denying it would be intellectually dishonest.

The problem is that productivity gains do not automatically become profit gains. In many industries, productivity improvements eventually reduce differentiation rather than increasing margins, because once every competitor gains access to the same acceleration capability, the competitive advantage begins to disappear. If every company can produce ten times more content, then content abundance itself becomes economically less valuable. If every development team can generate software dramatically faster, then the bottleneck shifts away from generation and toward validation, architecture, governance, operational stability, and trust.

AI Productivity Compression Risk

AI may be accelerating toward an outcome where knowledge production becomes abundant while economic differentiation becomes harder. The internet already demonstrated this dynamic: digital publishing massively increased content creation capability, but the abundance of content destroyed the economics of many traditional publishing models.

9. The Open Source Disruption Nobody Has Priced In

There is a second threat to the proprietary AI revenue model that receives far less attention than the capex sustainability question, and in some respects it is more structurally dangerous because it does not require a financing crisis to materialise. It is already here.

DeepSeek released its V4 model series as open source on 24 April 2026. The V4 Pro version targets high-performance use cases with overall capabilities approaching those of leading proprietary models. Both versions are available via an API that is drop-in compatible with OpenAI and Anthropic interface standards. That last detail deserves emphasis: an enterprise that has built its AI integration around OpenAI or Anthropic APIs can point those integrations at DeepSeek V4 with minimal code changes and run the workload on its own hardware. The switching cost that the land grab thesis depends on has been engineered around.

The enterprise case for local deployment of open source models is not primarily about cost, though the cost argument is real. It is about three things that regulated industries care about more than almost anything else: stability, compliance, and sovereignty.

On stability, a model running on your own infrastructure does not experience shared contention from millions of other users generating unpredictable latency. It does not get upgraded without your consent, breaking prompts and workflows that took months to calibrate. It does not go down when a hyperscaler has a network event precisely when your contact centre is handling peak volumes during a service incident.

On compliance, the fundamental architecture of a cloud API model means that data, prompts, and inference outputs traverse external networks regardless of enterprise privacy agreements. In banking, healthcare, and regulated financial services, the question is not whether the vendor’s data retention promises are credible. The question is whether the architecture is acceptable to the regulator at all, and in an increasing number of jurisdictions the answer is becoming no. Financial institutions are already using self-hosted open source models specifically to meet compliance requirements that cloud API deployment cannot satisfy, and the regulatory direction across GDPR, POPIA, and equivalent frameworks in Asia and the Middle East is consistently toward tighter data localisation. Every tightening of data sovereignty regulation is a tailwind for local deployment and a headwind for the per-token API model.

The open source AI market has grown 340% year on year in 2026, with enterprises deploying open weights models in production increasing from 23% to 67% of surveyed organisations. The strategic implication for the hyperscaler revenue model is sharp. If a material portion of production workloads migrate to locally hosted open source models, the addressable market for API inference contracts is structurally smaller than the current investment thesis assumes. The enterprises most likely to make that migration are exactly the ones that generate the highest volumes: regulated financial services, healthcare, and government.

10. The Historical Parallel Everyone Forgets

One of the easiest ways to dismiss criticism of AI economics is to compare the current moment to previous technological revolutions. Railways destroyed enormous amounts of capital before reshaping industrial civilisation. The internet vaporised countless companies before creating modern digital commerce. Cloud computing required years of infrastructure expansion before enterprise transformation stabilised economically. Those comparisons are valid, but they should create caution rather than comfort.

Historical Infrastructure Mania Comparisons

History shows that civilisation scale technological transitions often generate massive overinvestment phases where infrastructure is built ahead of sustainable monetisation. Some of that infrastructure becomes foundational for future economic growth, while enormous amounts of capital are simultaneously destroyed along the way. The existence of long term transformation does not guarantee short term profitability for the participants financing the transition.

The 1999 fibre overbuilding parallel is imperfect in its details but structurally recognisable: massive capital deployment, genuine long run utility, and a painful correction that sorts the infrastructure that found real demand from the infrastructure that did not.

11. What Does the Investment Unwind Look Like?

This is the question that most AI commentary declines to engage with directly, because answering it honestly requires specifying failure mechanisms rather than gesturing at systemic risk in the abstract.

OpenAI and Anthropic are two of the largest purchasers of AI infrastructure services from the hyperscalers. If either company fails to generate the revenue growth that justifies its compute contracts, the hyperscalers face revenue collection challenges on obligations they have already built infrastructure to fulfil. The hyperscalers have nearly doubled their collective weighting in the US investment grade corporate bond index over the year ending April 2026, issuing debt to fund infrastructure in advance of seeing returns. AI-linked issuance was roughly 5% of total investment grade corporate bond issuance in 2025, approximately three times the prior decade’s average annual technology issuance.

The unwind does not require a dramatic collapse. It requires a slower version of the demand growth that the build-out assumes. If open source local deployment captures even 20% of the enterprise production workload that would otherwise have run on hyperscaler APIs, the demand assumptions underwriting the infrastructure investment are materially wrong. If inference pricing continues to fall at the rates Epoch AI documents, the revenue per unit of compute falls even as volume grows. DeepSeek V4’s one million token context window and agent capability improvements are explicitly designed for the agentic workload that hyperscalers are counting on to drive the next consumption wave.

The most likely unwind scenario is not a sudden crash but a prolonged margin compression that forces explicit choices. Hyperscalers will face pressure from three directions simultaneously: rising depreciation from prior capex commitments, pricing pressure on inference revenue from open source alternatives, and demand shortfalls if the enterprise adoption curve is slower or more fragmented than the infrastructure build-out assumes. The response will be a combination of pricing discipline, where the below-cost subsidy era ends and inference prices rise toward economic cost; capex reduction, where future hardware orders are scaled back to match realised demand; and consolidation, where the model companies that cannot reach positive gross margin without the subsidy either find acquirers or fail. Oracle’s 57% stock decline since its OpenAI infrastructure pledge is an early indicator of what the market does when it loses confidence in a specific demand assumption.

For enterprises, the subsidised pricing era is a window of opportunity, not a permanent feature of the landscape. Organisations that use the subsidy window to build genuine capability, clean data foundations, and production deployments with demonstrated client value will be positioned to absorb the eventual repricing. Organisations that use it to run pilot programmes and maintain appearances will find that when pricing normalises, their AI programme becomes significantly more expensive and they have nothing built that justifies the increased cost.

12. The Enterprise Reality Is Far Messier Than the Demos

The gap between AI demonstrations and enterprise operational reality remains enormous. Demo environments are clean, constrained, and optimised for capability visibility. Real organisations are chaotic systems filled with legacy infrastructure, fragmented data quality, regulatory obligations, operational risk, inconsistent governance, and deeply human workflows.

Many organisations are already discovering that AI adoption creates entirely new categories of operational complexity. Generated code increases review demand. AI assisted support workflows introduce governance and liability concerns. Autonomous agents require extensive observability layers to avoid unpredictable behaviour. Compliance teams suddenly need explainability frameworks for decisions generated probabilistically rather than deterministically. Security teams now face entirely new attack surfaces through prompt injection, model abuse, and data leakage pathways.

Enterprise AI Reality Check

The share of companies abandoning the majority of their AI initiatives skyrocketed from 17% in 2024 to 42% in 2025, with 46% of all AI proofs of concept being scrapped before reaching production. The failure mode is not usually that the technology failed in isolation. It is that it could not scale across the business because the underlying data architecture was fragmented, the process owners had not changed their workflows, and the organisation lacked the connective tissue to convert a working demo into a production system. The 96% of financial institutions citing noisy or inaccurate data as their primary AI challenge are not experiencing a technology failure. They are experiencing a data governance failure they are trying to solve with an AI budget.

13. Pilots That Go Nowhere, and Why the Foundation Comes First

I can illustrate what this looks like from the inside with a specific example. Earlier this year Capitec launched Pulse, an AI system that surfaces a real time contextual picture of a client’s recent account activity to contact centre agents before the call is answered, reducing handling time by up to 18% and delivering a 26% net performance improvement across the pilot group. The coverage focused on the numbers. The interesting part is what had to exist before the numbers were possible.

Pulse was not built because Capitec wanted to demonstrate AI capability in the contact centre. It was built because we asked a specific question about a specific client: what does someone experience when they are frustrated enough to call us, and what would it take to make that moment effortless? A client who has had a transaction declined, or whose app has stopped working, or who cannot understand a charge on their account, is calling us at the worst moment in their relationship with the bank. The question was whether we could know why they were calling before they told us, and whether we could hand that context to the agent before the first word was exchanged.

That client outcome requirement drove every technical decision that followed. The data architecture had to be genuinely real time. The schema had to be clean and logically consistent. The read path had to be completely isolated from the write path.

Pulse does something that sounds straightforward: it reads live banking data across 25 million clients in the seconds between a client requesting a call and an agent answering it. Two obvious architectural approaches exist for any bank trying to replicate this. The first is to replicate all relevant data into a dedicated store and query the replica. The problem at banking scale is that replication lag under real load is measured in minutes not seconds for high-volume services, and during an outage, when contact centre demand peaks and the system is most needed, the replication overhead compounds precisely the pressure the production systems are already under. The second approach is to query the live production databases directly. The problem there is that at Capitec’s transaction volumes, adding a contact centre read workload directly onto the production write path generates measurable tail latency on payment processing. That is not a trade-off any bank can accept.

The only path to a working system was Aurora PostgreSQL with dedicated read replicas across the entire banking estate, where Aurora’s shared storage architecture means replica lag is measured in single digit milliseconds rather than minutes, where MVCC based read isolation means queries never acquire row level locks and therefore have no mechanism to interfere with the write path, and where all operational logs across the platform are coalesced into OpenSearch so that the system can correlate declined transactions, app diagnostics, service interruptions, and payment signals through a single indexed interface. This architecture was available to Capitec because Capitec owns the source code of its entire banking stack.

The AI layer, once the data problem was solved, introduced almost none of the architectural complexity of the preceding work. Claude reads the assembled context via Amazon Bedrock and produces a natural language briefing with a consistency that would have been implausible two years ago. The model required no fine tuning. It required good inputs. Every engineering decision described above is, in a precise sense, the work required to produce those good inputs reliably at speed.

Now consider what this means for a bank running a licensed third party core banking platform built on mainframe architecture. That organisation cannot make the Aurora PostgreSQL decision across its estate because it does not own the source code of its core banking systems. A vendor arriving with a Pulse equivalent product would encounter exactly the two failed options described above. The third option is not available to them regardless of budget or engineering capability, because it requires architectural decisions that had to be made years earlier before any specific AI use case was on anyone’s roadmap.

This is not a technology gap. It is a foundation gap. And it is the reason why the 42% of organisations abandoning AI projects are not primarily experiencing model failures or vendor failures. They are running into the limits of foundations they did not invest in before AI arrived.

14. The Structural Winners

The enterprise research consistently points to a small cohort of structural winners, and understanding what separates them from the majority is the most practically useful thing a senior technology leader can take from this entire body of evidence.

The first type are the infrastructure layer companies, positioned so deep in the AI supply chain that every dollar spent anywhere in the ecosystem flows through them regardless of which application or model wins. The valuation table in section 1 tells their story more compactly than any prose can.

Nvidia’s data centre revenue surged 75% year on year in Q4 FY2026 and reached a record $43 billion in the most recent quarter. The moat is not just the hardware. It is CUDA, a proprietary programming platform that developers have spent nearly two decades mastering, which means switching away from Nvidia is not a hardware decision, it is a retraining and re-engineering decision that most organisations cannot make at speed. Even the open source and local deployment trend offers Nvidia a partial hedge: DeepSeek V4 running locally on enterprise hardware is still running on Nvidia silicon.

TSMC’s full year 2025 revenue reached approximately $122.9 billion, representing 31.6% year on year growth, with net income of approximately $55.4 billion, the strongest annual performance in the company’s history. Every leading AI chip is fabricated at TSMC. Its leading edge nodes and CoWoS advanced packaging capacity are fully sold out, with Nvidia expected to consume more than 50% of CoWoS output through 2026.

ASML reported full year 2025 revenue of €32.7 billion with net income of €9.6 billion at a 52.8% gross margin, and its year end backlog stood at €38.8 billion providing multi-year revenue visibility. No other company on earth makes production ready EUV lithography machines. A single EUV system integrates over 120,000 high precision components and takes years to manufacture.

The second type of structural winner is the enterprise that has done the foundational work that makes AI deployable at scale. Companies identified as AI visionaries show 1.7x revenue growth and 3.6x three year total shareholder return versus laggards. Capital One is in its thirteenth year of technology transformation and ranked second on the 2025 Evident AI Index. JPMorgan has allocated nearly $20 billion to technology with over a thousand AI use cases running on a unified data platform. What these organisations share is not that they invested more heavily or moved faster than their peers. It is that they invested earlier in the preconditions. They built coherent data platforms before they needed them for AI. They redesigned processes rather than overlaying automation on broken workflows.

A study of 200 B2B AI deployments reinforces this point, finding that projects with smaller initial budgets achieved 2.1x higher ROI than large scale deployments, with a median return of 347% in year one and an eight month breakeven. The enterprise tendency to build programmes, hire consulting firms, and launch transformations is systematically destroying the thing that makes AI work: tight feedback loops, rapid iteration, and close proximity between the person with the problem and the person building the solution.

15. The Real Question: Can Civilisation Afford AI at Scale?

The most important question in AI is no longer whether the technology works, because clearly it does. The most important question is whether civilisation can economically sustain the infrastructure required to operationalise AI at planetary scale.

Every major AI breakthrough now immediately creates second order consequences across power grids, semiconductor manufacturing, datacenter construction, networking infrastructure, cooling systems, and geopolitical supply chains. That is not the behaviour of a lightweight software revolution. That is the behaviour of a civilisation scale industrial transition.

The AI Infrastructure Dependency Stack

The irony is that AI was initially marketed as the ultimate abstraction layer for knowledge work, yet the deeper the industry moves into large scale deployment, the more physical the economics become. Beneath every elegant AI demo sits an enormous stack of energy consumption, infrastructure financing, semiconductor dependency, and operational complexity. The question is no longer whether AI can generate intelligence. The question is whether the global economy can sustainably finance the industrial infrastructure required to make that intelligence universally available.

16. Why Companies Are Ignoring Profitability

Which brings us to the more interesting question. If the evidence is this clear, why does capital keep moving in this direction and why does the pace keep accelerating?

JPMorgan has characterised the dominant force behind current AI spending not as FOMO, fear of missing out, but as FOBO, Fear of Becoming Obsolete. The distinction matters. FOMO is opportunistic and assumes a prize worth chasing. FOBO is existential and assumes a penalty for inaction that compounds over time until it becomes irreversible. Organisations are not spending because they believe they can calculate the return. They are spending because not spending feels like a strategic abdication that cannot be explained to a board in three years if competitors have pulled ahead.

What is worth being direct about is that FOBO and genuine client value focus are structurally incompatible decision frameworks. An organisation spending because it fears obsolescence starts with external pressure and works inward toward a justification. An organisation spending because it has identified a specific client problem worth solving starts with a client and works outward toward whatever technology is required. The first framework produces ghost projects designed to demonstrate presence in the race. The second produces systems that make clients materially better off. Most enterprise AI programmes are being run under the first framework while being presented to boards as though they are the second.

The governance deficit compounds all of this. CEO involvement in AI investment decisions rose from 26% to 55% between 2024 and 2025, with CFO involvement going from 1% to 38%, yet only 33% of organisations report regular cross functional collaboration to align AI initiatives. More people are making the decisions and fewer people are collectively accountable for the outcomes. Most enterprises evaluate AI success through operational metrics such as efficiency gains and productivity improvements rather than financial outcomes, and these metrics cannot be traced to cost without analytical work that most organisations are not doing.

Citi’s identification of a 30 basis point credit spread penalty for companies classified as AI adopters versus AI enablers is a signal that the debt markets are beginning to price the distinction between activity and evidence, but that signal has not yet penetrated most boardroom conversations about AI strategy.

17. The Honest Conversation We Are Not Having

I presented Pulse to our Board earlier this year. The conversation that followed was entirely about clients. Which clients would benefit most. Whether the clients who most needed an effortless experience were the ones most likely to be reached by the system in its current form. What the handling time improvement meant for a client waiting on hold during a service interruption when they were already anxious about their money. The Board did not ask whether we were ahead of competitors on AI adoption. They did not ask which model we were using or what our AI strategy looked like relative to peer banks. They asked whether clients were better off, by how much, and how we would know.

That is the right conversation. It is also, based on everything in this article, the rarest one.

The question that changes the dynamic is not “is our AI profitable” in the abstract. It is “is this specific deployment making a specific client’s life materially better, at a cost we can justify, and how do we know?” That question cannot be answered with a vendor case study or a pilot group satisfaction score. It requires the same analytical discipline that any other capital allocation decision in a well-run organisation demands: a defined outcome, a measurement framework that tracks it financially, and a willingness to stop investing when the evidence does not support continuation.

None of this means the transformation is optional or the technology is overrated. The strategic gap between organisations that execute well and organisations that perform participation is going to widen significantly over the next three years regardless of what happens to hyperscaler financing or the open source disruption. The subsidy era of AI pricing will not last indefinitely, and the enterprises that have built genuine capability during it will be far better positioned to absorb the repricing than those who spent the same period accumulating pilots.

The companies that treat client value as the measure of success, and profitability as the financial proof of it, are going to be fine. The ones running AI programmes designed to demonstrate presence in the race rather than to win anything specific for the people they serve are funding someone else’s infrastructure build. History may eventually remember this era not simply as the birth of artificial intelligence, but as the largest infrastructure financing event humanity has ever attempted.

References

Ordered by importance to the central argument.

1. Is AI Profitable Yet? Industry Scorecard · Michael Tan Sikorski · May 2026
2. The Subprime AI Crisis Is Here · Ed Zitron, Where’s Your Ed At · March 2026
3. Capitec Pulse: The Engineering Behind Real-Time AI at Scale · Andrew Baker · March 2026
4. DeepSeek-V4 Preview Now Available with Open-Source Access · TechNode · April 2026
5. FOMO Has Proven a Stronger Incentive Than Poor Stock Performance · Fortune / Goldman Sachs · May 2026
6. $700 Billion in Capex. $50 Billion in Revenue. AI’s Math Is Broken. · Vin Vashishta · May 2026
7. AI ROI in 2026: Why Enterprise AI Fails and Works · Terminal X · April 2026
8. Beyond ROI: Are We Using the Wrong Metric in Measuring AI Success? · UC Berkeley SCET / MIT NANDA · February 2026
9. The AI Bubble: Hidden Risks and Opportunities · Man Group · May 2026
10. AI Capex Risk: Why AI Infrastructure Stocks Sold Off · HeyGoTrade · April 2026
11. The Price of AI: How Capex Is Rewriting Tech Balance Sheets · Breckinridge Capital Advisors · April 2026
12. FOBO and FOMO Supercharging AI Capex · Yahoo Finance / JPMorgan · 2026
13. Nvidia FY2026 Q4 Earnings · Nvidia SEC Filing · 2026
14. TSMC FY2025 Earnings and AI Chip Demand Analysis · TSPA Semiconductor · January 2026
15. ASML Q4 FY2025 Financial Results · ASML · January 2026
16. AI Capex 2026: The $690B Infrastructure Sprint · Futurum Group · February 2026
17. Hyperscaler Capex >$600bn in 2026: A 36% Increase · IEEE ComSoc · December 2025
18. LLM Inference Prices Have Fallen Rapidly · Epoch AI · 2025
19. AI Agents Burn 50x More Tokens Than Chats · LeanOps · May 2026
20. Microsoft Reports Are Exposing AI’s Real Cost Problem · Fortune / Gartner · May 2026
21. 2025 Forbes AI Study: Measuring Enterprise AI ROI · Mavvrik / Forbes Research · March 2026
22. Maximizing the Real ROI of AI in Operations · TechClass / IBM CEO study · April 2026
23. The Great AI Pullback: Evidence Why Enterprises Are Halting Pilots · Baytech Consulting · September 2025
24. AI ROI: Why Only 5% of Enterprises See Real Returns in 2026 · Master of Code · April 2026
25. DeepSeek and the Open Source AI Revolution in 2026 · Programming Helper · March 2026
26. Financial Institutions Using Self-Hosted Open Source LLMs for Compliance · ThinkPeak AI · November 2025
27. AI ROI Analysis: Evidence from 200 B2B Deployments 2022 to 2025 · Denis Atlan, ENDKOO · November 2025
28. Bank of America: Hyperscalers to Spend 90% of Operating Cash Flow on Capex · MarketWise · February 2026
29. Hyperscaler AI Spending Doubts Rising · Futuriom · April 2026