*Translated from zhichai.net. A scenario analysis dated August 10, 2026; not investment advice.*
Tonight I talked with a friend for a long time—from Alibaba's compute relationship with Moonshot AI, to ByteDance, DeepSeek, Zhipu, and MiniMax, then across the Pacific to the final fates of OpenAI, Anthropic, Google, and Microsoft.
By the end, the original simple question—"which model company will win"—had dissolved. The frontier model top spot is like a chair that keeps changing hands. What truly decides the endgame is what the leaderboards don't measure: who can secure compute sustainably, who retains core teams, who owns user entry points and enterprise workflows, who can bear the long capital expenditure, and who will be treated by state power as a strategic asset that cannot be allowed to fail.
> Model capability decides a battle; organization, infrastructure, and institutional position decide the length of the war.
Alibaba and Moonshot AI: Competitors Can Also Be Cloud Customers
According to a Bloomberg report relayed by Reuters, Moonshot AI has a compute arrangement with Alibaba allowing use of roughly 20,000 Nvidia chips. Reuters did not independently verify the report; what Alibaba explicitly denied was "providing H200 compute to Moonshot AI," not all compute cooperation. The cautious formulation: large-scale cooperation has high credibility, but chip model, contract value, signing date, deployment location, and settlement terms remain unconfirmed.
This is more interesting than "20,000 cards" itself. It reveals that model competition and cloud platform interests are not fully aligned. Kimi may compete with Qwen on some capabilities, but as long as it runs on Alibaba's infrastructure, Alibaba still gains cloud revenue and equity returns. Alibaba competes in the race while also owning the track.
This is Alibaba's deep competitiveness: cloud, model platforms, enterprise customers, open-source ecosystem, and consumer scenarios. Qwen doesn't need to be absolutely first in every generation for Alibaba to keep benefiting from AI industry expansion.
But infrastructure does not automatically produce next-generation models. Recent personnel and organizational changes around the Qwen core team remind us: the scarcest asset in frontier modeling isn't GPUs or released weights, but tacit knowledge accumulated by a group of people—data recipes, training anomaly judgment, memory of failed experiments, research taste, and collaboration chemistry. These cannot be handed over like servers.
So I don't think Alibaba has lost competitiveness, but I'd apply an "organizational continuity discount" to its model win rate. Its floor remains high; whether it raises its ceiling depends on the next flagship model fully led by the new team—not on how a launch event narrates change.
China's Commercial Winner May Be ByteDance
If I had to bet on China's overall AI commercial endgame, I'd currently favor ByteDance.
The reasoning isn't mysterious. Doubao, Douyin, CapCut, Feishu, and Volcano Engine form an extremely short feedback loop. Models quickly enter real products, reach massive users, and behavioral feedback returns to product iteration and inference infrastructure. ByteDance's strength has never been owning an isolated technology, but compressing technology into products ordinary people open repeatedly.
This makes the judgment uncomfortable: if AI becomes the total entry point for personal life and information distribution, the company that best understands attention, recommendation, and content consumption has a natural advantage. It may be most capable of winning—yet not the winner many would hope for.
I'd rather DeepSeek win.
DeepSeek may not end up with the largest revenue or the biggest entry point, but it has the best chance to change the industry's cost curve, open standards, and research direction. It's a rule-changer: approaching the frontier with fewer resources, diffusing influence through open weights, and forcing all competitors to recompute inference costs through low prices.
The "interview" with Liang Wenfeng circulating online is actually a closed-door investor briefing from May 20, 2026. The public version runs about 3 hours 44 minutes, transcribed by speech recognition and organized by AI; the original audio hasn't been released. The National Business Daily reported that participating investment institutions confirmed the meeting and its main content as credible. The reasonable way to use this material: trust its strategic skeleton, but don't treat every number or verbatim quote as audited fact.
What I see in this briefing is not an idealist rejecting commerce, but someone using commercial discipline to protect scientific ambition. His "restraint" is not a moral posture but resource prioritization: not letting short-term user scale, trendy multimodal businesses, or profit maximization prematurely rewrite the team's goals.
DeepSeek's real bet is a technical ladder: after chain-of-thought comes agents, after agents comes continuous learning, then models helping develop the next generation of models. If continuous learning is truly solved, research efficiency could compound, and DeepSeek could shift from "chasing with less compute" to "using intelligence to expand the productive capacity of intelligence."
But this remains a strong hypothesis, not a scheduled engineering task. If continuous learning stalls, or if AI's value ultimately settles in user entry points and cloud platforms, DeepSeek might gain enormous technical prestige while more economic value flows to ByteDance, Alibaba, and other platforms.
So China's endgame is better described not as winner-take-all, but as three victories coexisting: ByteDance claims the largest consumer entry point, Alibaba controls key cloud and enterprise infrastructure, and DeepSeek keeps changing model efficiency and the open ecosystem.
Zhipu and MiniMax walk narrower paths. Zhipu has the higher floor—a chance to become the strategic foundation-model supplier for government-enterprise, private deployment, domestic compute adaptation, and compliance systems. MiniMax has the higher ceiling—more like a global consumer AI company with in-house models, seeking breakthroughs in companionship, roleplay, voice, music, video, and multimodal creation. One is more stable; the other more volatile. Neither may become overall champion, but both could hold irreplaceable positions.
The American Endgame May No Longer Belong to Independent Labs
For the US, simply applying the internet-platform competition framework is no longer sufficient.
Advanced AI increasingly resembles the nuclear, aerospace, and defense industries: it requires massive chips, electricity, data centers, financing, government permits, and security review. As models enter defense, intelligence, research, and administrative systems, independent labs gradually lose their pure independence.
OpenAI is most likely to evolve into a "quasi-national lab and commercial contractor." The US government needn't legally acquire it; through procurement, defense contracts, security review, energy and data-center permitting, chip policy, and strategic financing, it can continuously shape OpenAI's boundaries.
This binding is both moat and cage. It raises OpenAI's survival priority, making it a strategic asset too important to fail, while eroding corporate autonomy and its neutral image in global markets. Measured by strategic power, OpenAI may become America's model-layer champion; measured by independence, it may be the most transformed.
Anthropic's main risk is not that Claude suddenly disappears, but that the company loses economic independence. High valuations, enormous compute costs, and dependence on Amazon and Google infrastructure and distribution could lead to acquisition, restructuring, or de facto control by a cloud platform even while the legal entity survives.
If the AI bubble bursts, Anthropic is more likely to face valuation resets, forced financing, and absorption of its technical assets than to see its research evaporate. The company may die, but Claude's technical lineage would enter a larger machine.
Google remains the full-stack player with the strongest recovery capacity: DeepMind, TPUs, Cloud, Search, Android, YouTube, and vast cash flow. But existing assets don't automatically generate the next technical route. Frontier modeling is a fluid capability tied to core teams and organizational consensus. Even with concerns about talent flow and model cadence, Google isn't out—it retains the most complete conditions for recovery, but must re-solve talent stability, organizational integration, and whether it dares to cannibalize traditional search advertising.
Microsoft may become the most stable economic value capturer. It doesn't need to own every strongest model; as long as Azure, GitHub, and Microsoft 365 carry government, enterprise, and developer workflows, OpenAI's nationalization, Anthropic's absorption by clouds, and the commoditization of model capability may all reinforce Microsoft's position as the "assembly platform."
The Real Winner May Be a Complex
China and the US may ultimately show different structures.
China is more likely to form a multipolar landscape of platforms, cloud, and independent research: ByteDance takes the biggest commercial entry point, Alibaba becomes key infrastructure, DeepSeek keeps the technology-disruptor seat. The US is more likely to form a "state–capital–cloud infrastructure complex": OpenAI deeply shaped by national strategy, Anthropic absorbed by cloud platforms, Google retaining a full-stack comeback option, and Microsoft extracting value from every outcome.
The real competitive unit of the future may no longer be a company, but an entire system converting energy, chips, capital, talent, institutions, and user entry points into continuous intelligence iteration.
Models will take turns at the top; applications will be replaced repeatedly; company names may be rewritten by acquisition and restructuring. But those who control compute, those who organize research communities, and those who hold distribution and government contracts will decide which power structure intelligence ultimately embeds in.
If tonight's discussion must be compressed into one sentence:
> ByteDance may claim China's AI commercial throne; DeepSeek is more likely to change what the throne is made of. OpenAI may become America's strategic champion—but those who end up with the most power and profit may not be any independent model lab at all.
---
*This is a scenario analysis as of August 10, 2026, and does not constitute investment advice. Sources: Reuters' relay of the Bloomberg report on the Moonshot–Alibaba compute cooperation, NBD's institutional verification of the Liang Wenfeng briefing, and Sina Tech's analysis of the transcript's evidentiary gaps.*
- Moonshot–Alibaba compute report: https://finance.yahoo.com/technology/ai/articles/moonshot-nvidia-chip-cluster-alibaba-150851345.html
- NBD verification of the 3h44m briefing: https://www.nbd.com.cn/articles/2026-07-23/4504599.html
- Sina Tech on transcript evidence boundaries: https://finance.sina.com.cn/tech/csj/2026-07-23/doc-iniiuwtf5389545.shtml