On October 28, 2025, in an internal memo distributed across Amazon's management ranks, Andy Jassy announced the elimination of fourteen thousand corporate positions, the largest corporate workforce reduction in the group's recent history. The public reading was immediate, almost mechanical: artificial intelligence is replacing humans. That reading is, on closer examination, misleading. To my mind, it errs both on the identity of the agent of substitution and on the nature of the economic dynamic it purports to describe.
At the time the memo circulated, no operational AI system at Amazon had automated the functions of the fourteen thousand people concerned. The majority occupied positions as project managers, recruiters, marketing leads or middle managers in support functions, whose internal processes remain, for the most part, manual. Jassy's move was therefore not a verdict on task obsolescence: it was a budgetary arbitrage. And it is precisely this distinction, between effective substitution and financing trade-off, that illuminates the sequence currently underway. The reality is less linear and more revealing: two distinct dynamics converge on the same result, a reduction in employment to the benefit of compute.
1. Big Tech: the preemptive investment strategy
Capital expenditure (CAPEX) indicators among the major technology firms point to an unprecedented acceleration. Amazon expects CAPEX exceeding $100 billion in 2025, a sharp increase over 2024. Microsoft reported an investment of $34.9 billion in the first quarter of fiscal 2026, an annual growth rate of roughly 74%, according to official earnings releases (8-K forms filed with the SEC). Google Cloud, for its part, anticipates $75 billion in infrastructure investment for 2025, a rise of more than 40% on the prior year.
It is crucial to note that these capital allocations are not principally aimed at automating existing internal processes. Their primary purpose is the acquisition of Nvidia graphics processing units (GPUs), the construction of data centers, and the build-out of large-scale artificial intelligence (AI) infrastructure. The strategy fits within a logic of preserving competitiveness over the medium and long term (a three- to five-year horizon). Brian Olsavsky, Amazon's CFO, has explicitly indicated that the fourth-quarter 2024 investment pace ($26.3 billion) would be representative of 2025, specifying that roughly 90% of that spending would be allocated to AWS, of which more than 90% directly to AI systems.
The backlog dilemma: when demand outstrips supply
A particularly revealing indicator of this dynamic is the AWS order book (remaining performance obligations), which stood at $195 billion in the second quarter of 2025, growing roughly 25% year on year. That figure, nearly twice AWS's annual revenue (around $110 billion), illustrates a situation of massive excess demand for compute power. Contracts are signed, revenues are secured, but delivery capacity is constrained by the inability to procure enough GPUs within the required timeframes.
The challenge is not unique to Amazon. Amy Hood, Microsoft's CFO, has reported capacity constraints ("we have run short of power and space") that froze close to $300 billion in unrecognized cloud contracts as of the time of writing (October 2025); that backlog has continued to grow in subsequent quarters, exceeding $600 billion according to Microsoft's statements in the third quarter of 2025. Likewise, Anat Ashkenazi, CFO of Google Cloud, has confirmed a similar situation of "more demand than available capacity" for a backlog of $106 billion.
It seems to me worth insisting on this point: the scarcity expressed here is not a scarcity of demand, nor even a scarcity of capital. It is a physical scarcity. Land, electrical interconnection, cooling water, high-bandwidth memory supply chains, advanced packaging capacity at TSMC. The hyperscalers are not saying they are short of ideas or customers, they are saying they are short of megawatts. That is a notable qualitative shift from previous cloud cycles, in which the dominant constraint was software optimization.
The strategic arbitrage: OPEX vs CAPEX
Faced with unmet demand and the need to sustain growth and margins, the hyperscalers are running a budgetary arbitrage. To support the explosion of their capex while preserving profitability, a drastic reduction in operating expenses (opex) becomes imperative. Payroll is invariably the largest line item within opex.
The 14,000 layoffs announced by Amazon in October 2025 (roughly 4% of corporate employees) are therefore not the consequence of AI automation already in place. They constitute a deliberate reallocation of capital: a reduction in current labor costs to finance the future technological infrastructure essential to honoring customer commitments and maintaining competitiveness. Microsoft followed a similar trajectory with 6,000 layoffs in May 2025, followed by 9,000 in July, even as the company committed to $80 billion in cloud and AI infrastructure investment for fiscal 2025. Meta, with its strategy of gradual 5% headcount cuts every six months, including 3,600 positions in January 2025 and a further 600 in the AI division in October 2025, illustrates a comparable approach.
It bears noting that the internal workflows of these organizations often remain characterized by manual processes. Current layoffs identify areas where a temporary reduction in human activity is judged acceptable in order to free up financial resources for GPU acquisition.
2. Traditional sectors: automation already at work
In parallel, companies in traditional sectors present a different picture. UPS announced over the course of 2025 several waves of headcount reductions (totaling, according to Reuters, up to 48,000 positions eliminated over the year across its network and support functions, with some internal tallies more modest). Nestlé announced in October 2025 a plan to cut 16,000 positions over two years (of which 12,000 white-collar), or roughly six percent of its global workforce. Ford, over the same interval, reduced its engineering and administrative headcount.
Here, the reason for the layoffs is not massive GPU investment but the maturity and effectiveness of AI automation already deployed. Salesforce is a telling example: its CEO, Marc Benioff, declared in 2025 that the firm's customer support headcount had fallen from 9,000 to 5,000 over the year, attributing this reduction of roughly 4,000 positions directly to Agentforce taking over about half of customer interactions. A significant share of the affected staff was redeployed to sales and other functions, according to Benioff's public statements (covered in particular by Fortune, CX Today and Salesforce Ben in 2025).
These firms have implemented operational AI solutions, such as conversational agents, predictive logistics-optimization systems, or generative design tools, producing measurable productivity gains. Their fundamental difference lies in the fact that they are not investing in the construction of massive GPU infrastructure, but rather leasing inference services (the execution of pre-trained AI models) from the hyperscalers. The economic calculus has tipped: the cost of automating certain tasks is now lower than the cost of the corresponding labor.
The Salesforce case carries, to my mind, the weight of a signal. When a vendor whose business was precisely to sell tools to human support teams publicly admits to having reduced its own human support by forty-four percent in twelve months thanks to its own agent, it validates the commercial argument it puts to its customers. For the finance directors of traditional companies, the move then becomes harder to dismiss. It is no longer a supplier's promise, it is a verifiable precedent.
3. The compute ecosystem: where value concentrates
These two dynamics, although distinct in their immediate causes, converge within a single economic ecosystem. The technology giants are building AI by investing in GPU infrastructure and data centers. Traditional companies consume those inference services and pre-trained solutions.
At the heart of this architecture sit the semiconductor manufacturers. TSMC produces the chips for Nvidia, which sells them to the hyperscalers at substantial margins. The hyperscalers train the models and provide the inference services that traditional companies exploit to automate. The result is a historic concentration of wealth in favor of TSMC, Nvidia and ASML, while employment simultaneously declines at both ends of the value chain: the tech firms lay off to finance infrastructure, and traditional firms lay off because that infrastructure is generating gains. Value therefore crystallizes in the compute layer (chips, hardware) rather than in the human layer.
AI startups: explosive growth, dizzying deficits
The AI startup ecosystem presents an intriguing paradox: meteoric revenue growth coexists with abyssal losses. OpenAI recorded, according to data reported by Sacra and corroborated by several sources, around $4.3 billion in revenue in the first half of 2025, more than the entirety of 2024, and was then targeting an annual figure of around $13 billion. ChatGPT reaches 700 million weekly users.
Yet these numbers are eclipsed by colossal deficits. According to the figures reported by Microsoft in connection with its stake in OpenAI, the lab is said to have incurred losses of several billion dollars per quarter, with R&D spending far exceeding revenues. Anthropic, whose valuation followed a meteoric trajectory in 2025 (on the order of sixty billion dollars in mid-year, multiplied several times over in the following months), then reported projected revenues in the billions of dollars after raising a massive cumulative investment total. Perplexity AI, over the same period, saw its valuation triple in a few months to settle around nine billion dollars.
These massive investments, principally directed toward GPU acquisition, reflect a strong anticipation of future returns, despite the current unprofitability of the models.
A structural shortage, not a speculative bubble
The hypothesis of a speculative bubble is often advanced. The facts, however, contradict that reading. A bubble would be characterized by excess supply confronting insufficient demand. Yet the order books of AWS ($195bn), Microsoft ($300bn) and Google Cloud ($106bn) reveal massive and unmet demand for AI services. The CFO statements ("short of power and space," "strong demand, limited supply") confirm a structural shortage. The situation indicates that firms perceive substantial value in these investments.
4. The counter-argument of the financeable backlog
A serious objection deserves to be confronted head-on. If aggregate demand for compute is as massive as the backlogs suggest ($195, $300, $106 billion of signed but undelivered contracts at the three hyperscalers), then the layoffs are not the confession of a crisis but a margin optimization in a period of growth. The debt is financeable. The contracts are secured. Shareholders applaud.
The objection holds in the short term. It holds less well, it seems to me, along two dimensions that market triumphalism tends to erase.
First dimension: the visibility of backlogs. A remaining performance obligation (RPO) does not unwind mechanically into revenue. A substantial share of these commitments stretches over several years, and their recognition depends on the operational ability of the hyperscalers to deliver the promised compute. Microsoft, Google and AWS have publicly acknowledged in 2025 a shortage of power, land and GPUs. If delivery extends over five to seven years rather than three, the present value of these contracts declines, and the pricing power of the suppliers in the lower layer (Nvidia, ASML, TSMC) increases commensurately.
Second dimension: the pricing power of the upstream layers. The H100, then the H200, then Blackwell have seen their unit prices rise with each generation, in an environment in which Nvidia's gross margin sits at historically elevated levels. The share of margin that flows to the final customer, that is, to the deployer of AI applications, is by construction the residual of this chain. For a traditional company that adopts an AI agent whose bill is itself indexed to the inference cost of OpenAI or Anthropic, themselves indexed to the marginal compute cost from Nvidia, productivity rises, but a growing share of the gain flows back to Taiwanese silicon.
The counter-argument of the financeable backlog therefore describes the hyperscalers' situation faithfully; it does not describe the situation of employees, nor that of downstream customer companies. It is precisely this dissociation between the health of compute and the health of work that the next section labels a prosperous depression. To my mind, treating this dissociation as a bookkeeping detail rather than as a structuring economic fact is the principal blind spot of market commentary in 2025.
The opening phase of a profound transformation
Statistics on enterprise AI adoption suggest we are at the beginning of this transition. According to the successive surveys of the U.S. Census Bureau (Business Trends and Outlook Survey) and of McKinsey (State of AI 2024-2025), enterprise adoption rose from a very low base in 2023 to about ten percent in 2025 for deep integration, while a majority of organizations worldwide now report using AI in at least one function. Generative AI in particular saw its adoption more than double between 2023 and 2024.
Historically, and according to the work of Carlota Perez and Brian Arthur on technological waves, the most disruptive phase of a revolution occurs when adoption spreads from the ten-percent threshold toward the majority of the market. We sit, for deep AI integration into business processes, at the beginning of that phase, which suggests sustained growth for the next three to five years.
The paradox of a prosperous depression
The current situation presents a destabilizing paradox. For financial markets and shareholders, the context is favorable: strong growth, rising valuations, optimistic outlook. Within firms, however, the reality is different. Constraints on payroll are strict, employment anxiety is intensifying, and the workload of the remaining staff is rising, leading to a deterioration of morale. The scenario evokes, to my mind, a "prosperous depression": prosperous for capital, but depressive for labor.
The expression is no rhetorical flourish. It describes a precise macroeconomic configuration in which the aggregate indicators (market capitalizations, margins, productivity per head) improve in step with the lived indicators (a sense of instability, work intensification, the rarefaction of internal promotions, freezes on junior hiring). The two series do not contradict one another, they simply describe two distinct actors in the same accounting equation. When payroll falls by five percent and the value produced remains stable, the ratio improves mechanically for the shareholder and deteriorates just as mechanically for the remaining employee, who is implicitly asked to absorb the difference.
Conclusion: a nuanced reality with identical consequences
The reality of AI's impact on employment is more complex than a binary simplification allows. While the complete automation of complex roles is not yet widespread (Amazon cannot attribute 14,000 layoffs to operational AI), AI already excels at automating specific and repetitive tasks, which accounts for the workforce reductions at UPS (48,000) or Salesforce (4,000).
Current layoffs therefore stem from two distinct dynamics. At the technology giants, they fall under a strategic reallocation of capital, sacrificing human jobs to finance future GPU infrastructure; today's wage costs pay for tomorrow's silicon. At traditional companies, they translate, by contrast, into a workforce reduction directly tied to measurable productivity gains generated by AI automation already in operation, that is, a substitution effect already observed on specific tasks, not a bet on infrastructure to come.
These two vectors converge on a massive redistribution of capital, away from wages and toward computing infrastructure. The dynamic enriches chipmakers disproportionately, while employment declines on both fronts. Demand for AI is substantial and growing, far from a speculative bubble. It represents a complex transformation that affects different firms for different reasons, yet invariably arrives at the same result: a reduction in human jobs, an increase in the machine share, and a growing concentration of wealth in the domain of compute at the expense of labor.
This diagnosis aligns with what the OECD documents in its Compendium of Productivity Indicators 2025 under the name of the superstar firms phenomenon: a small number of very large companies capture a disproportionate share of productivity gains, invest massively in information technologies, and concentrate the bulk of innovation. Artificial intelligence does not create this phenomenon, it amplifies it. I have developed this point in a parallel essay, What We Replace, What Resists, which attends more specifically to the redistribution of tasks within occupations and to the mapping of what resists automation.
At a glance
The AI war in numbers: two distinct stories, one single outcome. On one side Big Tech, on the other traditional companies, but the same dynamic of capital reallocation, from human labor toward compute.
Story 1: Big Tech
2025 layoffs
| Company | Jobs cut | Detail |
|---|---|---|
| Amazon | 14,000 | 4% of 350K office employees |
| Microsoft | 15,000 | 6K in May + 9K in July (3% of workforce) |
| Meta | 4,200 | 3.6K in January + 600 in October (5% every 6 months) |
| Big Tech total | ~33K |
2025 CAPEX investments
| Company | 2025 CAPEX | Detail |
|---|---|---|
| Amazon | $100B | vs $83B in 2024 (+20%), 90% for AWS |
| Microsoft | $80B | $34.9B in Q1 FY26 (+74% year over year) |
| Google Cloud | $75B | vs $52.5B in 2024 (+43%) |
| Total CAPEX | $255B |
Cloud backlog (signed contracts not yet recognized)
| Player | Backlog | Detail |
|---|---|---|
| Amazon AWS | $195B | +25% year over year, i.e. 2x AWS annual revenue |
| Microsoft Azure | $300B | Unrecognized contracts |
| Google Cloud | $106B | $58B over 2 years |
| Total backlog | $601B |
A brutal reallocation: turning OPEX (salaries) into CAPEX (GPUs).
Story 2: traditional companies
2025 layoffs
| Company | Jobs cut | Detail |
|---|---|---|
| UPS | 48,000 | 10% of 490K employees (34K operations + 14K managers) |
| Nestle | 16,000 | 12K white-collar + 4K production/logistics (over 2 years) |
| Salesforce | 4,000 | Customer support: 9K to 5K (-44%) |
| Ford | ~3,000 | Engineering and administration 2024-2025 |
| Traditional total | ~71K |
Real automation
Salesforce already automates 50% of its customer conversations with AI. The concrete levers:
- AI conversational agents
- Supply chain optimization
- Generative design tools
Business model
- Renting cloud inference services
- Pre-trained models (AWS / Azure / GCP)
- Turnkey solutions
Equation: cost of automation < cost of salaries. Productivity gains that are real, measurable and documented.
Total 2025 layoffs
~104,000 jobs, tech and traditional companies combined.
The big winner: the compute layer
| Link | Player | Position |
|---|---|---|
| GPU | NVIDIA | Margins of 60%+ |
| Chips | TSMC | Exclusive manufacturer |
| Machines | ASML | EUV monopoly |
Record-high historical margins while employment collapses on both sides. Wealth concentrates in the compute layer, never in the human layer.
AI startups: wild growth, abyssal losses
OpenAI
| Metric | Value |
|---|---|
| H1 2025 revenue | $4.3B |
| H1 2025 losses | -$13.5B |
| H1 2025 R&D | $6.7B |
| Weekly users | 700M |
| Revenue targets | $13B in 2025, $100B in 2027 |
Anthropic
| Metric | Value |
|---|---|
| Annualized revenue | $1.4B |
| Growth | +40% |
| Total raised | $18.5B |
| Valuation | $58 to $61.5B |
| 2025 projection | $2 to $3.7B |
Perplexity AI
Sharply rising valuation: $3B in June 2024.