The AI Spending Spree is Evolving

Image source | Internet (please contact for deletion if there is any infringement). Partially AI-generated.

For the past two to three years, conversations about major AI companies have revolved around one key theme: profligate spending.

Those with the deepest pockets and the boldest strategies seemed to hold all the cards at the high-stakes poker table.

Trillion-dollar infrastructure blueprints were unveiled one after another. Data centers sprang up across the landscape, and GPUs were purchased in vast quantities.

This was an era where ambition was gauged by the scale of investment, following a straightforward and unforgiving logic: everyone wanted a seat at the table first, figuring out the details later.

But now, subtle yet profound shifts are underway.

Microsoft is trimming budgets for external model usage and steering internal AI tools toward its proprietary systems. Meta, while maintaining multi-billion-dollar infrastructure investments, is implementing layoffs affecting around 8,000 roles to 'mitigate AI-related cost pressures.'

In China, Alibaba, Tencent, and Baidu continue to see significant year-on-year increases in capital expenditures. However, the capital market's response post-earnings announcements has been notably consistent: it's less concerned with the amount invested and more focused on the returns generated.

These signals collectively point to a clear conclusion: the AI race hasn't shifted from 'high spending' to 'low spending.' Rather, it has transitioned from 'reckless spending' to 'responsible spending.'

The competitive landscape remains unchanged; what has evolved are the evaluation criteria and the behavior of the participants.

Spending money is no longer a guarantee of success; what matters is how that money is spent.

Microsoft and Meta Hit the Brakes

Many observers interpret terms like 'cuts' and 'contraction' as signs that major players are exiting the race. However, a closer look at their specific actions reveals a different story.

Microsoft's most notable move has been to slash the budget for utilizing Anthropic's technology (primarily Claude Code) by more than one-third from the anticipated minimum of $1 billion, while encouraging employees to adopt more in-house AI tools.

At Meta, the number of employees using Claude Code has dropped from approximately 60,000 at the start of the year to around 30,000, with users of self-developed coding tools now exceeding that number.

These companies are not 'abandoning AI' but rather 'ceasing to pay indiscriminately for external models.'

The financial rationale is straightforward. Meta spent over $105 million on Claude Code in just 28 days. As AI tools transition from niche innovations to ubiquitous production tools, costs based on user seats and usage volumes can skyrocket.

Meanwhile, the performance of in-house models is improving, and the cost of building and operating inference systems is declining.

Switching from external procurement to internal development is a logical step for any technologically advanced company at this juncture.

What's even more telling is the shift in the capital market's focus. Goldman Sachs stated bluntly in a mid-October research report that the market's attention on Microsoft is shifting from the scale of AI investment and revenue growth to the unit economics of infrastructure and return on invested capital (ROIC).

Microsoft extended the useful life of some data centers from 15 to 25 years and revised its capital expenditure forecast for the 2026 calendar year from approximately $190 billion to about $175 billion. These adjustments essentially address the same question: how much sustainable revenue is generated for each dollar invested?

Meta's situation further illustrates this point. Its capital expenditure guidance for 2026 remains as high as $125 billion to $145 billion, a significant increase from $72.2 billion in 2025. The absolute scale of investment is not shrinking.

However, management has publicly acknowledged 'uncertainty' about the widespread success of AI consumer products before committing to such heavy investments.

This sends a clear message: when investment scales reach a certain level, 'daring to spend' is no longer a matter of courage but of financial prudence.

Reallocating resources from layoffs to compute investments, which Goldman Sachs refers to as 'replacing cages with birds,' is not merely 'budget-cutting for the winter.'

The cage is being upgraded, not dismantled.

A Financial Equation That Doesn't Add Up

If this were merely a pacing adjustment by individual companies, it wouldn't signal the end of the era of competitive spending.

What truly renders this path unsustainable is a structural financial gap.

Gartner's October forecast presents a stark contrast: global AI infrastructure spending is expected to approach $1 trillion in 2026, with an additional $6 trillion to be invested between 2027 and 2030.

However, by 2030, AI data center owners will need to generate a cumulative $8.5 trillion in revenue to cover capital expenditures, financing costs, and achieve a 12% ROIC.

Yet, the same forecast suggests that enterprise and consumer AI revenue attributable to data centers during this period will reach a maximum of only $3.6 trillion.

The gap is not marginal but more than double. Gartner warns that if current capital expenditure trends continue, over $3.5 trillion in AI infrastructure investments could face impairment risks by the end of 2030, and a 12% ROIC may not be achievable until the end of 2032 at the earliest.

This data exposes the fundamental contradiction of the 'spending race': the growth rate of compute supply is systematically outpacing the rate of revenue generation.

Over the past two years, major players could justify investments with a 'build first, figure it out later' approach because demand was indeed surging, and supply was indeed insufficient.

But as supply gradually catches up and even exceeds short-term demand in some areas, if the growth rate of capital expenditures continues to outpace that of operating cash flow, the financing environment will shift from 'supportive' to 'restrictive.'

A Columbia Business School professor, whose insights have been cited by multiple media outlets, predicts that in 2026, the capital expenditures of several ultra-large platforms will exceed their combined operating cash flow for the first time.

This means the room for funding unlimited expansion with internal cash flow is narrowing.

What truly enables 'rationalization' are concurrent technological advancements.

Inference costs have plummeted over the past year, with some infrastructure service providers reporting drops of up to 90% and projecting further declines.

NVIDIA's new-generation system has reduced the inference cost per million tokens to as low as 1/45th of the previous generation's product. When 'doing more with less compute' becomes a reality, efficiency becomes a more critical competitive factor than scale.

Advantages once achieved through sheer compute power can now be obtained through model optimization, system scheduling, and software-hardware collaboration to achieve higher output from the same resources.

Spending remains crucial, but the marginal returns on spending are declining, while the marginal returns on efficiency are rising.

The Investment Paradigm Has Shifted

Shifting focus back to China, a seemingly contradictory phenomenon emerges amid the 'rationalization' narrative: BAT's capital expenditures are still growing rapidly.

In the second quarter of 2026, Alibaba's capital expenditures reached 67.678 billion yuan, up 75% year-on-year; Tencent's were 52.8 billion yuan, up 176%; and Baidu's were 11.4 billion yuan, up 201%. On the surface, this remains a high-investment posture.

However, the financial structure and strategic intent behind these figures have substantially diverged from the 'indiscriminate spending' phase.

Alibaba's approach is particularly illustrative. It merged its Cloud Intelligence Group and T-Head into 'AI Cloud & Computing Services' and reclassified its model labs and Qianwen-related businesses under 'AI Labs & Applications.' The intent is clear: to showcase the separate ROI for infrastructure and application layers, rather than letting e-commerce profits obscure AI's true progress.

During the same period, the adjusted operating profit of AI Cloud & Computing Services reached 5.628 billion yuan, up 133% year-on-year, with the profit margin rising to 12%.

This demonstrates that while compute expansion is still a drag on the group's cash flow, the cloud business itself is beginning to show operating leverage.

Although Alibaba continues to spend, it is simultaneously demanding that certain business units prove their ability to generate their own cash flow.

Baidu's changes are more akin to a 'strategic pivot' than a 'scaling up.' In the second quarter of 2026, Baidu's core AI-driven business revenue reached 12.5 billion yuan, accounting for half of its general business revenue, with AI cloud infrastructure revenue at 7.3 billion yuan, up 50% year-on-year.

However, total revenue was still declining, with traditional advertising contracting faster than AI revenue growth, and net profit attributable to the parent company declining significantly year-on-year.

Baidu's choice is no longer to compete on model parameters in general-purpose large models but to shift its focus to intelligent agents, with dense rollouts of product lines such as general-purpose agents, code generation, and decision-making agents, using 'daily active intelligent agents' rather than token consumption as the new value metric.

This is a classic example of substituting scale-for-scale competition with differentiated positioning: concentrating resources on application layers with shorter commercialization paths when compute capacity is not dominant.

Tencent's strategy is to leverage its robust cash flow to provide tolerance space for AI investments while clearly distinguishing between 'short-cycle investments' and 'long-cycle investments.'

Management's statement on the earnings call was straightforward: using GPUs for ad technology is a short-cycle investment that can quickly boost click-through rates and revenue, whereas large model training offers no direct returns in the short term and requires a full lifecycle perspective.

On the surface, this is not 'abandoning long-term investments' but placing long-term investments within a time-disciplined framework, rather than indefinitely postponing all return questions with 'long-termism.'

ByteDance is another noteworthy case. Its AI capital expenditures are reportedly at the 200 billion yuan level, but the money is highly concentrated in directions directly integrated with its content ecosystem, such as video models, Coding, Agents, and Doubao monetization.

Fanqie Novels provides IP, video generation models handle visualization, Douyin and short drama platforms manage distribution, and Volcano Engine handles compute and model scheduling.

This forms a closed loop from content production to distribution, where AI is no longer a 'second curve' hovering above the business but an accelerator embedded within the existing flywheel.

When investments are tightly integrated with the existing ecosystem, the return path for each dollar becomes shorter and more traceable.

The differences between China and overseas are less about 'who is more rational' than about differing constraints and pacing of rationalization.

Overseas giants, with their deeper cash flows, can afford to bet earlier and more aggressively, which is why the adjustment signals become clearer and more public when return gaps emerge.

Domestic players remain aggressive in absolute capital expenditure growth, but financial pressures such as declining profits and tightening free cash flow are forcing them to enter 'financial validation' mode earlier while still investing.

Although the paths differ, the direction is consistent: shifting from 'who spends the most' to 'who spends the most accurately.'

Spending Now Demands Efficiency

Synthesizing these trends, a more nuanced judgment emerges: the era of competitive spending has not been 'ended,' as AI infrastructure still requires massive investments, and Gartner's trillion-dollar spending forecasts will not disappear due to pacing adjustments by a few companies.

What has truly ended is the notion that the scale of capital expenditure alone constitutes competitiveness.

The next phase of competition will be reordered along at least three dimensions.

First, unit economic efficiency will replace total compute capacity as the key metric. When inference costs drop by 90% in a year and single-token costs are compressed to a fraction of their predecessors, the competitive focus shifts from 'how many cards you have' to 'how many effective tokens each card can produce and how much revenue each token can generate.'

Microsoft's focus on ROIC and Alibaba's separate disclosure of operating profits for its AI cloud business are moves in the same direction. Compute capacity itself does not constitute a barrier; the conversion efficiency of compute does.

Secondly, ecosystem integration plays a pivotal role in determining the duration of the return cycle. Apple's re-evaluation within the AI narrative is precisely attributed to its decision to abstain from the colossal, hundred-billion-dollar infrastructure race, which involves AI expenditures in the billions. Instead, Apple leverages its existing hardware and ecosystems for monetization.

This is not to suggest that Apple's approach can be easily replicated; rather, it highlights that the pace of AI investment returns is contingent not on the absolute investment amount but on the proximity of those investments to existing revenue streams.

ByteDance's seamless integration of AI into its content distribution ecosystem, Tencent's prioritization of GPUs for advertising technology, and Alibaba's incorporation of models into cloud and enterprise service contracts all effectively bridge this gap.

Thirdly, financial discipline is emerging as a competitive edge. As the industry transitions from a focus on 'growth narratives' to 'delivery narratives,' companies capable of sustaining longer investment cycles with healthier cash flow structures will maintain their momentum, while others may be compelled to scale back.

Meta's strategic use of layoffs to mitigate infrastructure costs, Microsoft's adjustments to depreciation policies and asset structures to regulate spending, and domestic companies' rigorous stress testing of profits and cash flow amidst surging capital expenditures all share a common thread: they are setting boundaries for 'sustainable spending' rather than viewing 'spending' as an end in itself.

In truth, the era of reckless spending has not vanished; it has merely ceased to be the ultimate objective.

The significance of the 2026 inflection point lies not in the extent to which certain companies have slashed their budgets but in the fundamental shift in the industry's evaluation criteria.

Previously, the question was, "Do you have the courage to invest?" Now, the inquiry has evolved to, "When will your investments become self-sustaining?"

As AI continues to mature, the competition has entered a more demanding phase: it is no longer about who is bolder but about who demonstrates greater excellence.