NextFin News -- Even as the slogan “cut costs and boost efficiency” still echoes across big tech companies’ internal networks, the capital expenditure figures on their financial statements have already delivered the answer.

In Q2 2026, Alibaba and Tencent together spent more than RMB 120 billion in capex. Tencent stood out most: despite posting annual profits of over RMB 100 billion, the giant reported free cash flow of -RMB 13.8 billion—its first time ever turning negative. Tencent’s capex payments in Q2 were about RMB 59.3 billion, of which roughly RMB 51.4 billion was prepayments for computing power procurement. Once paid, these amounts immediately show up as cash outflows, and the -RMB 13.8 billion figure was largely driven by that prepayment.

Other players are stepping on the gas as well. Baidu’s single-quarter capex reached RMB 11.4 billion, up about 200% year on year; ByteDance raised its full-year 2026 capex plan to RMB 200 billion, the most aggressive spending pace among domestic peers.

This is no ordinary capex cycle. China’s internet giants are locked in an arms race for a ticket into the AI era, with control of the next decade at stake.

Four Tech Giants, Four Paths to Survival

Hardware procurement makes up the bulk of capex. Of Alibaba’s roughly RMB 126 billion in capex for fiscal year 2026, AI servers, data centers, and GPU purchases were the unquestioned main event. Computing costs are still climbing: memory-chip prices have surged, and NVIDIA recently raised product prices by more than 15%, directly pushing up procurement costs for Chinese buyers.

Domestic substitution is penetrating at an unprecedented pace. By industry-survey estimates, Tencent’s share of domestic-chip procurement is expected to rise to more than 65% going forward, with training cards sourced mainly from Enflame and inference cards mainly from Cambricon. Under the sword of Damocles of geopolitics, this is both a “spare tire” being put into regular service and a defensive counterstrike.

The endgame of compute is electricity. Alibaba recently completed a HKD 80 billion placement—its first since its Hong Kong listing in 2019—with all proceeds to be invested in AI infrastructure. Sovereign wealth funds across the Middle East, Europe, and Asia subscribed for more than 40% in total. Long-horizon capital that has been tested through cycles worldwide is putting real money behind China’s AI narrative.

Focusing on these four companies themselves, they are not following the same playbook; instead, they have evolved four starkly different rules for commercial survival.Alibaba is the quintessential “water seller,” and it also has the clearest full-stack closed loop: selling computing power to enterprise customers. Q2 revenue came in at RMB 269.0 billion, up 9% year on year, but adjusted net profit was only RMB 20.7 billion, down 38% year on year—profits were visibly eroded by AI investment. The payoff is just as straightforward: revenue from AI cloud and compute services reached RMB 48.4 billion, up 45% year on year, while EBITA profit hit RMB 5.6 billion, up 133% year on year. With its public-cloud foundation, Alibaba is currently the fastest at converting AI spending into real revenue.

In contrast stands Tencent. Tencent has internalized massive computing power into WeChat’s ecosystem of 1.3 billion users, effectively playing the role of a “sub-landlord.” Q2 revenue surpassed RMB 200.0 billion for the first time, posting RMB 204.8 billion, up 11% year on year, but capital expenditures of RMB 52.8 billion far exceeded market expectations. The core issue with this model is that computing power is consumed by AI applications within WeChat; the payback cycle is long and opaque—hence the negative free cash flow of RMB 13.8 billion.

Baidu is the smallest by scale. Among these companies, it relied most heavily on monetizing through “selling compute and tokens,” and its first-mover advantage once delivered an initial wave of dividends. But earnings quality was under significant pressure: Q2 net profit attributable to shareholders was only about RMB 2.3 billion, down 68% year on year. With Alibaba Cloud and Volcano Engine suppressing it through sheer scale, how long Baidu can hold this position will depend on whether it can narrow the gap with top-tier cloud providers.

The most aggressive investor is ByteDance. Its planned RMB 200.0 billion in capital expenditures ranked first among the four, with Volcano Engine advancing on two fronts—internal demand for compute and external service delivery. As an unlisted company, ByteDance did not face the capital market’s quarterly profit pressure, so its spending wasn’t constrained by near-term results; however, whether Volcano Engine can truly scale up, and whether its business model can deliver, remains the most uncertain piece of the puzzle among the four.

China–U.S. AI Race Is About Tempo, Not Just Returns

Investment that spares no expense ultimately has to stand the test of the capital markets. In a global frame of reference, Goldman Sachs’ 2026 AI investment map shows that total worldwide AI-related investment will top US$1 trillion, with nearly US$600 billion in the U.S. alone—close to 60% of the total. But scale by itself doesn’t explain everything; what matters more is how structural differences play out. In the U.S., hyperscale cloud providers dominate: all three follow highly similar playbooks, and their capex cycles are almost perfectly in sync. In China, by contrast, the landscape is extremely fragmented: four models are placing bets at the same time—those selling compute, those internalizing ecosystems, those monetizing via tokens, and those investing purely for strategic reasons. Fragmented, and moving fast.

The gap in AI capital returns between China and the U.S. is, at its core, a divide between two stages of industry development. Morgan Stanley estimates that AI investment ROIC for North American cloud providers is around 25%–50%, versus just 13%–20% for their Chinese counterparts. The gap is real, but it needs to be assessed within each market’s development timeline: U.S. cloud providers began deploying AI infrastructure at scale around 2020 and have now entered a “harvest phase,” with compute leasing and large-model API calls generating steady cash flow while the marginal cost of expansion keeps falling. Chinese players’ large-scale spending has been concentrated in 2025–2026—effectively compressing what took the U.S. three years of build-out into a single year. With upfront investment amortized more intensively, a temporarily lower ROIC follows the logic of the industry cycle and does not, in itself, signal a competitiveness problem.

On structural advantages, China is not at a disadvantage. On the manufacturing side, the same amount of capex can deploy more physical compute, with a clear cost edge. On the applications side, China hosts the world’s largest markets for the industrial internet, autonomous driving, and smart cities—monetization is far from limited to selling compute. On the energy and supporting-infrastructure side, the pace of renewable-grid integration, UHV transmission rollout, and nuclear build-out is faster than in the U.S., which could translate into a better long-run operating-cost curve.

Goldman Sachs’ August report projected that in 2026, U.S. AI investment would add only about 0.1 percentage point to GDP growth on a net basis, with marginal benefits already declining. Soochow Securities estimated that AI would contribute roughly 0.3–0.5 percentage point to China’s GDP growth, with that boost not yet fully realized. As compute assets continue to depreciate and the application ecosystem matures, the ROIC of China’s AI investment is likely to improve gradually over the next 18 to 24 months. The China–U.S. AI race isn’t about who’s better—it’s about tempo.

Three Things That Decide Life or Death

This trillion-yuan-scale mega bet will ultimately be decided by three variables.

The first is tied to Nvidia’s Rubin cycle. On August 26, Nvidia released its latest earnings report: Q2 revenue hit $96.2 billion, up 106% year on year. Vera Rubin had already entered mass production, with Q4 shipments ramping up; price hikes of more than 15% were set to take effect when deliveries begin in early 2027, passing straight through to procurement costs in China. Faced with surging costs, China’s tech giants will have to choose between continuing to take on orders and accelerating a pivot to homegrown chips.

The second is the inflection point for domestic substitution. Tencent’s more-than-65% share of domestically sourced chip purchases is a bellwether. Only if Enflame and Cambricon can scale output and yields on schedule will Morgan Stanley’s modelled bind—where “higher hardware costs” drag down ROIC—have a chance of being fundamentally reversed. This is the key to lowering costs, and it’s also the dividing line.

The third hangs over the consumer market. For now, on the revenue side, it’s essentially all about selling shovels and water to businesses. Whether a ChatGPT-level, truly breakout product can emerge in China will directly determine whether Tencent’s internalization approach and ByteDance’s investment-led approach can actually deliver. Without an explosion on the consumer side, excess compute will eventually face a reckoning.

Evaluating this AI race purely through a rational ROI framework leads to a misread of its underlying logic. At its core, this is a survival game driven by the fear that “if you don’t invest, you’re out.” The HK$80 billion placement drawing subscriptions from sovereign wealth funds across multiple regions already says it plainly: even the most cautious long-term capital has begun to vote with its feet.

For these giants, the real question давно stopped being “does the math add up,” and became “if it doesn’t add up for us—but it does for our rivals—what then?”

The gap between Chinese cloud providers’ ROIC of just 13%–20% and the 25%–50% posted by their U.S. peers still has no definitive answer as to how it will ultimately converge: will it be leveled by a surge in domestic chip capacity, or will it be forced into a shakeout by a bubble bursting? Time is running out for the giants.