Before the bill comes due: reading enterprise AI through a FinOps lens

Executive summary

Enterprise AI spending is being committed today against two assumptions, and both are weakening.

The first assumption is that the price you see is the price you will pay. It is not. The major AI vendors are selling access below cost, and they are now starting to say so out loud. GitHub has announced that Copilot moves to token-based billing on 1 June 2026, telling customers the old flat-rate model “is no longer sustainable”. Anthropic briefly removed Claude Code from its USD 20 Pro plan in April 2026 before reversing within a day. These are early signals from a sector, not isolated pricing tweaks: a sector that, on its own numbers, is spending roughly USD 1.69 for every dollar it earns.

The second assumption is that a headcount decision made now can be unwound later if it proves wrong. It can be, but not cheaply and not quickly. In 2025, US employers attributed 54,836 announced job cuts to AI. A growing share of the companies that made those cuts are already regretting them: Orgvue found that 55% of leaders who made roles redundant because of AI now say they got it wrong, and Gartner predicts that half of all companies that cut customer service staff because of AI will be rehiring for the same work by 2027.

This paper does not argue that enterprises should slow down on AI. It argues that they should price the decision honestly. The discipline to do that already exists. Cloud cost practitioners spent close to twenty years watching cloud follow exactly this arc: subsidised pricing, enthusiastic adoption, structural lock-in, and then a repricing that arrived only after the switching costs were sunk. Cloud is the dress rehearsal. AI is the same play, performed faster, and most enterprises are buying tickets before they have read the script.

What follows is the evidence, the precedent, and the questions a CFO and CIO should be asking now, while the answers still change the outcome.

The precedent FinOps already lived through

Enterprise AI pricing is heading where cloud has already been.

For close to twenty years, the working assumption inside every IT planning function was that cloud costs only move in one direction: down. That assumption was earned. Since Amazon launched EC2 in 2006, AWS alone has made more than 100 price cuts across its services [1]. Competitive pressure between AWS, Azure, and Google Cloud kept compute and storage on a steady downward path, and a generation of infrastructure decisions was made on the expectation that the trend would continue.

In 2025 and 2026, that trend broke.

Two changes make the break concrete rather than anecdotal. First, AWS raised prices on EC2 Capacity Blocks for ML, the resource enterprises depend on for AI training and inference workloads, by roughly 15% in early 2026 [1]. Second, and more quietly, Microsoft removed the Enterprise Agreement volume discount tiers for its online services. In an announcement dated 12 August 2025, effective 1 November 2025, Microsoft eliminated EA discount Levels B through D for online services [2]. Independent analysis from Directions on Microsoft put the practical effect plainly: volume customers had typically received discounts of 6% to 12%, so the largest enterprise buyers, those at Level D with 15,000 or more users, lost the full 12% [3]. List prices did not change. The effective price went up by 6% to 12% for exactly the customers least able to switch.

That is the pattern worth holding onto. The repricing did not announce itself as a price increase. It arrived as a structural change, after adoption was complete and switching costs were sunk.

Here is the part that should concern any CFO funding an AI programme. Even with a mature, well-resourced cost discipline in place, the cloud is still not under control. Flexera’s 2025 State of the Cloud report, drawn from 759 cloud decision-makers, estimated that 27% of cloud spend is wasted, that organisations exceed their cloud budgets by 17%, and that 84% cite managing cloud spend as their single biggest cloud challenge [4]. Those are the numbers after close to twenty years of cloud cost practice, a mature tooling market, and seven years of a dedicated FinOps profession.

And the figure is now moving the wrong way. Flexera’s 2026 report records wasted cloud spend ticking up to 29%, reversing a five-year downward trend, and attributes the reversal specifically to AI workloads [5]. The FinOps Foundation’s own data shows why: the share of practitioners managing AI spend jumped from 31% to 63% in a single year, the largest mover in the 2025 State of FinOps report’s rankings, and reached near-universal adoption in the 2026 follow-up [6]. AI did not just join the cloud bill. It started driving the waste.

This is the precedent. Cloud was sold cheap, adopted broadly, locked in structurally, and then repriced. Mature cost discipline reduced the damage but never eliminated it, and AI is now pushing the waste line back up. The enterprises that fared best were the ones that treated cloud as a cost to be governed from day one, not a utility to be trusted. The next section shows the AI vendors running the same sequence, only faster, and largely before any equivalent governance exists.

The same playbook, in fast-forward

The cloud repricing took close to twenty years to arrive. The AI repricing has started inside three.

On 27 April 2026, GitHub announced that every Copilot plan would move to usage-based billing on 1 June 2026 [7]. The mechanics matter. Premium request units are being replaced by GitHub AI Credits, and usage is now metered on token consumption, counting input, output, and cached tokens at each model’s API rate. GitHub’s own explanation of why is the most useful sentence in the announcement: “a quick chat question and a multi-hour autonomous coding session can cost the user the same amount. GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable” [7]. The base subscription price did not change. What changed is that the meter is now running, the cheaper-model fallback is gone, and heavy users will pay materially more. One customer on GitHub’s own discussion forum estimated their cost moving from roughly EUR 67 in April to a projected EUR 966 a month under the new model. Treat that as an illustration rather than a benchmark, but the direction is not in dispute.

Anthropic ran a smaller version of the same test. Overnight on 21 April 2026, its pricing page changed to show Claude Code as no longer included in the USD 20 Pro plan. The company’s head of growth, Amol Avasare, confirmed to The Register that “we’re running a small test on ~2 percent of new prosumer signups”, and the change was reversed within 24 hours after public backlash [8]. The reversal is less interesting than the reason the test happened at all. The Register’s assessment was blunt: “Anthropic’s subscription plans charge far less than the book value of tokens consumed, sometimes by a factor of ten or more” [8]. A pricing test that gets reversed in a day is not a failed decision. It is a vendor checking how much tension the rope can take.

Neither of these is an isolated commercial choice. They are symptoms of an unsolved unit-economics problem that runs the length of the industry.

The clearest framing of the scale of that problem comes from David Cahn, a partner at Sequoia Capital. In “AI’s $600B Question”, published in June 2024, Cahn estimated the annual end-user revenue the sector would need to generate to justify its data-centre build-out [9]. His method is simple enough to audit: take Nvidia’s run-rate revenue, double it to cover the total cost of ownership of a data centre because the GPUs are only about half the bill, then double it again to cover the 50% gross margin a cloud provider needs to make reselling that compute. The result was a roughly USD 600bn annual revenue requirement against realised AI revenue running at a fraction of it. Cahn called it a USD 500bn hole that still needs to be filled. He is a venture investor in the sector. He has no incentive to talk it down, which is exactly what makes the number worth taking seriously.

The hole is visible in the accounts of the largest model provider. According to financial documents reported by The Wall Street Journal and summarised by Fortune in November 2025, OpenAI expects roughly USD 22bn in 2025 spending against about USD 13bn in revenue, which works out to spending approximately USD 1.69 for every dollar of revenue it generates [10]. The company’s own projections, reported by The Information, imply losses rising as high as USD 14bn in 2026, close to triple the prior year, with no profit expected until 2029 [11]. For context, OpenAI’s 2024 numbers, a roughly USD 5bn loss on roughly USD 3.7bn of revenue, already implied spending of about USD 1.35 for every dollar earned. The loss per dollar widens as the company scales rather than narrowing.

This is the part a CFO should sit with. A vendor losing USD 1.69 for every dollar of revenue has only three ways out: raise prices, cut the cost of inference faster than usage grows, or find external funding indefinitely. The first two land directly on the customer’s bill. The third is not under the customer’s control. GitHub and Anthropic are simply the first to start adjusting the first lever in public.

The harder question is whether the underlying compute will get cheap enough, fast enough, to make the current pricing sustainable without a customer-facing increase. The most prominent sceptic on that point is not a short-seller but Goldman Sachs’s own head of global equity research, Jim Covello. In Goldman’s June 2024 “Top of Mind” report, Covello argued that AI would need to “solve complex problems” to justify its cost, something he said the technology “isn’t built to do”, and explicitly compared the current capex cycle to “virtual reality, the metaverse, and blockchain” [12]. A CFO does not have to agree with Covello to take the point that matters here: the people closest to the money are not treating the favourable pricing as durable. Neither should the people signing the contracts.

The human-capital decision is harder to reverse than to make

The pricing risk in the previous section is recoverable. A contract can be renegotiated, a vendor can be switched, a workload can be moved. The decisions covered in this section are different in kind, because they involve people, and people do not switch back on like a service.

Start with the scale. In 2025, US employers announced 1,206,374 job cuts, up 58% on the prior year, the highest annual total since 2020 [13]. Within that, Challenger, Gray & Christmas, the long-standing industry tracker, attributed 54,836 announced cuts specifically to AI [13]. The trend did not slow into 2026. In March 2026, AI was the single largest stated reason for job cuts of any category, accounting for 15,341 of the 60,620 cuts announced that month [14]. Company by company, the attribution is explicit: Amazon cut roughly 14,000 corporate roles, Microsoft about 15,000 across the year, and Salesforce eliminated around 4,000 customer support roles, with each company naming AI as a factor [15].

The relevant question for a CFO is how many of those cuts were correct rather than whether they happened, and the early evidence is not reassuring.

Orgvue, an organisational-design software firm, surveyed 1,163 senior leaders in 2025. Of the 39% who had made employees redundant as a result of deploying AI, 55% said they had made the wrong decision [16]. More than a third reported employees quitting as a direct consequence of the AI rollout [16]. That is the people who made the cuts telling a surveyor, within months, that they got it wrong. Orgvue’s chief executive, Oliver Shaw, put the underlying error plainly: dismissing employees “without a clear plan for workforce transformation is reckless” [17].

Klarna is the case study that makes the abstraction concrete. The payments company spent 2024 as the public face of AI-driven headcount reduction. In February 2024 it announced that its AI assistant was handling the workload of 700 customer service agents. Its full-time headcount fell from 5,527 at the end of 2022 to 3,422 at the end of 2024, a reduction of roughly 38% that the company itself attributed to AI in its March 2025 IPO prospectus [18]. Then it reversed. By May 2025, chief executive Sebastian Siemiatkowski told Bloomberg the company was hiring human customer service agents again, because the AI-only model had produced lower-quality support. His stated reasoning is worth quoting in full, because it is a recantation from the executive who built the original case: “From a brand perspective, a company perspective, I just think it’s so critical that you are clear to your customer that there will always be a human if you want” [19].

What Klarna learned in private, Gartner has now turned into a forward prediction. In February 2026, Gartner forecast that by 2027, 50% of companies that cut customer service staff because of AI will rehire for similar functions, “but under different job titles” [20]. The prediction is signed by a named analyst, Kathy Ross, and based on a survey of 321 customer service and support leaders. Read it as a planning input, not a slogan: an established analyst firm is telling its enterprise clients to expect that half of this specific category of decision will be unwound within roughly a year.

Here is the asymmetry a CFO should price. Cutting a role is fast, clean, and shows up in the next quarter’s numbers. Rebuilding the capability is slow, expensive, and shows up nowhere flattering. The cost is not limited to recruiting and onboarding the replacement. It lands on the people who stayed. Research into “survivor syndrome”, pioneered by Columbia Business School’s Joel Brockner and studied for close to 40 years, finds that employees who keep their jobs after a downsizing consistently develop anger, fear, distrust, and guilt, and that these reactions reduce productivity and creativity while raising absenteeism and voluntary turnover [21]. That is a measurable drag, and it arrives precisely when the organisation needs its remaining people most.

The contemporary data lines up with the academic tradition. A 2025 Careerminds survey of 600 respondents found that 47% reported productivity declines following job cuts and 41% reported further talent leaving voluntarily afterwards [22]. The same survey found that nearly half of organisations rehired up to a quarter of the roles they had cut within a year [22]. A reduction partly reversed within 12 months is a costly round trip rather than a saving, with morale damage and lost institutional knowledge charged along the way.

None of this argues against ever reducing headcount. It argues that a headcount decision justified by AI capability should be underwritten with the same rigour as the AI capability itself. Which raises the question the next section addresses directly: how good is the evidence that the AI actually does the work?

A prior question: is the AI delivering?

The two preceding sections assumed the AI works. That assumption deserves its own scrutiny, because the most rigorous evidence available is not flattering, and a CFO underwriting either a vendor commitment or a headcount cut is implicitly underwriting it.

Start with the cleanest study on the question. In July 2025, METR, a non-profit AI research organisation, published a randomised controlled trial measuring how early-2025 AI tools affected the productivity of 16 experienced open-source developers across 246 real tasks in codebases they knew well [23]. The developers expected AI to make them 24% faster. After the study, they estimated it had made them 20% faster. Measured against the control condition, they were actually 19% slower, with a confidence interval running from 2% to 39% slower [23]. The slowdown is not the part that should worry a CFO. What should is that the practitioners doing the work could not tell. They believed they had been sped up by 20% while the stopwatch said the opposite. Self-reported productivity gain, the metric most AI business cases quietly rely on, was not merely imprecise here. It pointed the wrong way.

Honesty requires the update, and it cuts in more than one direction. METR ran a second study from August 2025 and reported in February 2026 that it could no longer measure the effect cleanly [24]. The reason is itself informative: between 30% and 50% of developers declined to submit tasks they did not want to attempt without AI, and others refused to participate at all rather than work unaided. The raw 2026 numbers still showed a slowdown, an estimated 18% for returning developers and 4% for newly recruited ones, but METR judged these a lower bound, called the data “only very weak evidence”, and concluded the true 2026 effect is probably better than 2025 while declining to attach a reliable figure to it [24]. A careful reader should hold both halves of that. The tools are now wanted enough by their users that a rigorous productivity experiment has become hard to staff, which is real evidence of perceived value. And yet METR, having watched its own measurement break, still would not claim a measured net gain. The defensible position in May 2026 is that AI coding tools are improving and are genuinely valued by the people using them, and that no rigorous study has yet demonstrated the net productivity gain that adoption decisions assume.

What holds for one developer and one task also holds at the level of the enterprise programme. In August 2025, MIT’s NANDA initiative published “The GenAI Divide”, which found that 95% of enterprise generative-AI pilots delivered no measurable profit-and-loss impact [25]. That headline has been criticised, and fairly: “success” was defined narrowly as measurable P&L impact within a short window, and the figure rests on a limited interview base. But two of the study’s secondary findings survive the caveat and are more useful than the headline anyway. Pilots built on tools bought from specialised vendors succeeded roughly 67% of the time, while pilots built in-house succeeded roughly a third of the time [25]. The failure mode, in MIT’s analysis, was integration rather than model quality: generic tools that do not adapt to a specific workflow stall once they meet a real enterprise process.

Gartner, working from a poll of more than 3,400 organisations, put a number on where that leads. It predicts that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls [26]. Its analyst, Anushree Verma, also flagged “agent washing”, the rebranding of older chatbots and automation software as agents, and estimated that only about 130 of the thousands of vendors claiming agentic capability genuinely offer it [26]. A CFO comparing vendor proposals should read that as a direct warning about what is actually being purchased.

None of this says AI does not work. It says the evidence base for the productivity gains being capitalised into vendor contracts and headcount plans is thinner, and more contested, than the pace of those decisions implies. Goldman Sachs’s Jim Covello, revisiting his 2024 scepticism in May 2026, reached a similar conclusion from the macro data: the spending has not yet produced a commensurate, measurable return [27]. For a CFO, the operative word is “yet”. It may. The point is that “it will” is currently a forecast, not a finding, and a forecast belongs in a sensitivity analysis, not in a committed cost base.

The regulatory and concentration overhang

Two further risks sit outside the unit economics but bear directly on the cost of an AI commitment: what compliance will cost, and how few suppliers the market actually has.

On compliance cost, the honest answer is that even the estimate is contested. The EU AI Act’s high-risk requirements include, among other things, a Quality Management System. The study commissioned to support the Commission’s impact assessment, conducted by CEPS with ICF and Wavestone, estimated that setting up a new QMS would cost between EUR 193,000 and EUR 330,000 up front, plus roughly EUR 71,400 a year to maintain [28]. Those costs fall only on providers of high-risk AI systems, and fully only on organisations that do not already run a QMS. The larger numbers that circulate widely, a EUR 400,000 per-SME burden and a EUR 30bn-plus hit to the EU economy, come from a separate Center for Data Innovation paper [29]. CEPS, the body that produced the underlying study, has publicly rebutted both: the EUR 400,000 figure, it says, was reached simply by adding its own EUR 330,000 setup and EUR 71,400 maintenance estimates together, and the EUR 30bn figure wrongly applied a high-risk compliance overhead to all AI investment rather than to the roughly 10% of systems that are actually high-risk [28]. The point for a CFO is that the compliance line is real, material, and still disputed at the level of its primary sources, which is the worst combination to budget against.

The timeline carries its own hazard, and it is not the one most plans assume. The AI Act’s high-risk obligations were due to apply from 2 August 2026. On 7 May 2026, EU negotiators reached a provisional agreement, part of the “Digital Omnibus” simplification package, to defer them: standalone high-risk systems under Annex III move to 2 December 2027, and high-risk systems embedded in regulated products move to 2 August 2028 [30]. That agreement still requires formal adoption. The lesson is that the regulatory date is itself a moving variable, and a compliance programme scoped to a fixed 2026 deadline was scoped to a number that has already changed once and could change again.

The second risk is structural. The frontier-model market is not a broad competitive field. Writing in The Guardian in September 2025, the investor Roger McNamee counted five or six essentially identical large-language-model efforts, each chasing the same global position [31]. The dependency is tighter than even that small number suggests, because the suppliers are financially entangled with one another. OpenAI’s February 2026 funding round raised USD 110bn at a USD 730bn pre-money valuation, with USD 50bn from Amazon and USD 30bn each from Nvidia and SoftBank; the round was later reported to have closed larger still [32][33]. A significant share of that backing was committed not as cash but as cloud and chip services, and OpenAI in turn committed to consuming Amazon and Nvidia compute [32]. An enterprise signing an AI contract is therefore not simply buying from a vendor. It is buying into a small, capital-intensive, circularly financed cluster of suppliers, where one participant repricing or retrenching would ripple straight through to the customer. That is concentration risk, and it belongs on the same register as the pricing and human-capital exposures already described.

What this means for the CFO and CIO

This paper has deliberately avoided telling anyone to adopt or abandon AI. That decision depends on a business case this document cannot see. What it can offer is a set of questions a CFO and CIO should be able to answer before the relevant cheques are signed, because each one maps directly to a risk the preceding sections have evidenced.

The first is whether AI spend is being governed with the discipline already applied to cloud. The FinOps profession exists because untracked, unallocated, optimistically forecast cloud spend still wasted 27% to 29% of the budget after close to twenty years of cloud cost practice. AI spend is currently being committed with far less of that machinery in place, and on pricing the vendors themselves are signalling will not hold. The practical test is simple: can the organisation attribute AI cost to a team, a product, and a unit of output today? If it cannot, the discipline needs to exist before the volume scales, not after.

The second is whether contracts are written for the pricing the vendors are moving toward rather than the pricing on offer now. GitHub’s shift to token billing and Anthropic’s Pro-plan test are the visible edge of a sector spending well above its revenue. A contract underwritten on today’s flat or subsidised rate, with no price-protection clause, no modelled usage-based scenario, and no tested exit or portability path, is a contract underwritten on an assumption the counterparties are actively withdrawing.

The third question, and the most important, is whether any headcount decision attributed to AI could be reversed if the capability underperforms, and what that reversal would cost. The evidence in this paper is that this category of decision is regretted often, reversed within roughly a year at material expense, and damaging to the people who remain. A headcount reduction justified by AI should clear a higher evidential bar than a reversible decision would, and the business case should carry the rehiring cost as a quantified downside scenario rather than an afterthought.

The fourth is whether the AI capability underpinning all of the above has been measured rather than assumed. Self-reported productivity gain, the metric most business cases quietly lean on, ran in the wrong direction in the most rigorous study available. A pilot that defines success after the fact, measures against no control, and trusts the enthusiasm of its users is not evidence. It is a forecast wearing the costume of one.

None of this is an argument against AI. It is an argument for sequencing. Cloud rewarded the organisations that built cost governance early and penalised those that bolted it on after the bill had already scaled. AI is running the same arc on a shorter clock: subsidised pricing, fast adoption, structural lock-in, and a repricing that, on the vendors’ own statements, has already begun. The contracts and the headcount decisions being made now are being priced against the most favourable conditions the market is likely to offer. The discipline that cloud cost management spent close to twenty years learning is available off the shelf. The advantage goes to whoever applies it before the bill comes due, not after.

Sources

1. Amplix, What AWS GPU Pricing Reveals About Cloud Cost Risk, 2026.
2. Microsoft, Online Services Pricing Consistency Update, 12 August 2025.
3. Directions on Microsoft (Mary Jo Foley), analysis of the August 2025 EA online-services pricing change.
4. Flexera, 2025 State of the Cloud Report (14th annual; survey of 759 cloud decision-makers), 19 March 2025.
5. Flexera, 2026 State of the Cloud Report.
6. FinOps Foundation, The State of FinOps 2025 Report (861 respondents, roughly USD 69bn in cloud spend), February 2025, with 2026 follow-up data.
7. GitHub, GitHub Copilot is moving to usage-based billing, The GitHub Blog, 27 April 2026.
8. The Register, Anthropic tests reaction to yanking Claude Code from Pro, 22 April 2026.
9. David Cahn, “AI’s $600B Question”, Sequoia Capital, 20 June 2024.
10. Fortune, OpenAI says it plans to report stunning annual losses through 2028, 12 November 2025 (citing Wall Street Journal financial documents).
11. The Information, OpenAI Projections Imply Losses Tripling to $14 Billion in 2026 (paywalled).
12. Goldman Sachs Research, Gen AI: Too Much Spend, Too Little Benefit?, Top of Mind, Issue 129, 27 June 2024.
13. Challenger, Gray & Christmas, Inc., 2025 Year-End Challenger Report: Highest Q4 Layoffs Since 2008; Lowest YTD Hiring Since 2010, January 2026.
14. Challenger, Gray & Christmas, Inc., Challenger Report: March Cuts Rise 25% From February, AI Leads Reasons, April 2026.
15. CNBC, AI was behind over 50,000 layoffs in 2025, here are the top firms to cite it for job cuts, 21 December 2025.
16. Orgvue, 55% of businesses admit wrong decisions in making employees redundant when bringing AI into the workforce (survey of 1,163 senior leaders, conducted by Vitreous World), 29 April 2025.
17. TechRepublic, ‘We Acted Too Quickly’: Over Half of Companies Regret AI-Driven Layoffs, Report Finds, 20 May 2025.
18. Payments Dive, Klarna whittled workforce via AI ahead of IPO (citing Klarna’s March 2025 IPO prospectus).
19. Entrepreneur, Klarna Is Hiring Customer Service Agents After AI Couldn’t Cut It on Calls, According to the Company’s CEO (quoting Sebastian Siemiatkowski’s May 2025 Bloomberg interview).
20. Gartner, Inc., Gartner Predicts Half of Companies That Cut Customer Service Staff Due to AI Will Rehire by 2027, 3 February 2026.
21. Andreas Werr and Wiley Wakeman, Dealing with “survivor syndrome”, Stockholm School of Economics (summarising the survivor-syndrome research tradition pioneered by Joel Brockner, Columbia Business School).
22. HR Dive, Most HR leaders conducted “serial layoffs” within the past year, per survey (reporting a 2025 Careerminds survey of 600 respondents), 15 August 2025.
23. Joel Becker, Nate Rush, Beth Barnes, and David Rein, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, METR, 10 July 2025 (full paper: arxiv.org/abs/2507.09089).
24. METR, We are Changing our Developer Productivity Experiment Design, 24 February 2026.
25. Fortune, MIT report: 95% of generative AI pilots at companies are failing (reporting MIT NANDA, “The GenAI Divide: State of AI in Business 2025”), 18 August 2025.
26. Gartner, Inc., Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025.
27. Fortune, “FOMO has proven a stronger incentive than poor stock performance”: Goldman Sachs finds insecurity is a key part of the AI boom, 6 May 2026.
28. Moritz Laurer, Andrea Renda, and Timothy Yeung, Clarifying the costs for the EU’s AI Act, CEPS, 24 September 2021.
29. Benjamin Mueller, How Much Will the Artificial Intelligence Act Cost Europe?, Center for Data Innovation, July 2021.
30. Council of the EU, Artificial Intelligence: Council and Parliament agree to simplify and streamline rules, press release, 7 May 2026.
31. Roger McNamee, AI investors are in for a rude awakening, The Guardian, 24 September 2025.
32. CNBC, OpenAI announces $110 billion funding round with backing from Amazon, Nvidia, SoftBank, 27 February 2026.
33. CNBC, OpenAI closes record-breaking $122 billion funding round as anticipation builds for IPO, 31 March 2026.