62. Year-End AI Review: What Emerged from Interviews with More Than 30 AI Practitioners in China and the US?

Summary

• The US-China AI narratives formally diverged this year: US capital assigned OpenAI and its peers an AGI premium akin to sovereign-grade assets, while Chinese assets entered the “read the financials, DAU and retention” phase earlier. Cheapness is the surface; discount is the substance—“the combined valuation of all Chinese AI startups may not even match Thinking Machines Lab’s debut valuation,” so dollar VCs are backing Chinese teams for engineering, applications and Agents, not the foundation-model dream.
• Reverse CFIUS (the OIR) has pushed Chinese AI founders into a stark era of choosing sides, while Benchmark, Sutter Hill, Greylock and other top-dollar institutions have tacitly accepted a “decoupling of capital and talent” framework. Benchmark’s $75M investment in Manus followed by questions from the US Treasury was the defining event; HeyGen bought back early shareholders and moved its headquarters to Los Angeles, successfully threading the extremely narrow path of “Chinese founders + China’s technology dividend + Silicon Valley capital + global markets,” while lawyer 黄敏达 cautioned that Manus’s move to Singapore “is not worth emulating.” Any Chinese company raising more than $100M will face the choice of cutting itself off from China.
• China’s primary market enjoyed a brief spring, with capital flowing into four major areas—embodied intelligence, AI applications and Agents, multimodality, and AI hardware—and hardware attracted more funding than software in Q2 2025. By Q2, total funding into embodied intelligence and AI hardware had exceeded RMB14.5B, with leading embodied-AI companies holding at least RMB1B in cash; China’s smart-glasses market grew 121% year over year, and AI glasses are seen as the key battleground for the next-generation interface because they occupy the face’s “core triangle.”
• 云启’s 陈宇 argues that the financing window for top-tier projects is often only a few months, much like autonomous driving in 2016: the companies still at the table today all emerged within 1-2 quarters back then. The speed at which non-consensus becomes consensus has compressed sharply because six or seven years of technical groundwork had already accumulated before ChatGPT’s breakout—the missing ingredient was capital; but “there are very few institutions and people in China’s market capable of making independent judgments,” and most projects are merely “reading newspapers on an iPad; Toutiao has not appeared yet.”
• Silicon Valley’s leaderboard shifted rapidly this year: ChatGPT was still challenging for SOTA at the start of the year, Anthropic led at year-end, and Gemini took over in the second half; after GPT-5, “the ceiling of the scaling law under the existing Transformer architecture may already have appeared.” Google’s Gemini 2.5 Pro became the first model ever to top the text, vision and Web-development rankings simultaneously, while Nano Banana restored the company’s standing, putting intense pressure on OpenAI and pushing it toward de-Microsoftization through Stargate; Oracle, the “super infrastructure contractor brought in with its own money,” surged nearly 40% in a single session before falling nearly 40% from its high, with free cash flow turning negative.
• Meta’s multibillion-dollar acquisition of Manus is a bet on Agent momentum: perhaps even without building the strongest model, building the strongest Agent execution layer could still produce an entirely different outcome. Many Silicon Valley practitioners and observers believe Meta has “fallen behind across the board” on the foundation-model front and will remain behind for at least another six months; four AI teams are operating in parallel internally, Yann LeCun has left, and short sellers have returned to the secondary market—across the major tech companies, the problem “is not innovation, but organization.”
• The scaling law remains valid, but its form is changing: from scaling model size to scaling inference-time thinking depth, and next to scaling the network effects of Agent collaboration, while “the low-hanging fruit may soon be gone.” The wall-builders—LeCun and Ilya—disagree with the non-believers-in-death—Hinton and Dario; a Google DeepMind researcher’s boundary thesis is that “language is the boundary of human knowledge, but not the boundary of knowledge in the universe,” and “AGI can approach but cannot arrive.”
• The commercial reality is harsh: OpenAI has $20B in annualized revenue, but card-rental costs alone run to $16B and operating losses to $11.92B—every dollar earned requires $1.6 of investment. “If the compute consumed in training is a drop of water, inference may consume an entire pond”; Chinese foundation models have commercially overtaken through open source and cost-performance, with Zhipu priced at one-seventh of Anthropic and winning sovereign contracts in multiple countries, while the pragmatic path for startups is model orchestration—“combining the strengths of multiple models.”

Deep dive

1. DeepSeek’s First Aftershock: The Bamboo Ceiling Vanishes from AI

• Shijie laid out the methodology behind the review: it draws on nearly 30 practitioners from China and the US and hundreds of thousands of Chinese characters of recordings. Because “it is hard to find a guest with sufficiently focused and accurate observations on every direction,” the team cast a wide net across people and questions to find cross-validation across fields while preserving disagreements.
• Richard, head of 大观资本’s North America business, relayed the reaction of his white friend at Nvidia after reading DeepSeek’s source code: “This engineering is beautiful.” The deeper reason Chinese people became AI’s main workforce is that Chinese engineers built much of Silicon Valley’s early infra—“infra is grueling; white people don’t want to do it”—and infra happens to be critical to foundation models.
• Several dollar investors told Shijie that once Chinese projects such as DeepSeek, Qwen and K2 appeared, “the bamboo ceiling in Silicon Valley AI had already disappeared entirely.”

2. The Second Aftershock: From “No Point in Blocking” to “Not Blocking Enough”

• After DeepSeek R1 produced a GPT-4o-class model with less compute and greater efficiency, the US split into two camps. One, represented by Nvidia, argued that the compute chokehold had been proven ineffective and should be abandoned; the other, represented by Anthropic CEO Dario, argued that “the problem is not that we blocked too much for nothing, but that we did not block enough.”
• The second camp expanded rapidly after Trump took office. Tighter compute controls, stronger investment restrictions and the AI Diffusion Rule together pushed the US-China technology rivalry to a new peak, providing the year’s overarching macro backdrop.

3. Manus and Benchmark’s Super Check: Confidence Ignites

• Manus exploded in March, was labeled a “second DeepSeek” by Chinese state media, went through investment scrutiny midyear and was acquired by Meta for billions of dollars at year-end. In Q1, from DeepSeek and Unitree to Manus, “the entire Silicon Valley was talking about China,” and market sentiment shifted.
• Benchmark’s founder came to China personally to hunt for projects, then invested $75M in Manus parent company Butterfly Effect a few months later, pushing its valuation to $500M. The amount was not large by Silicon Valley standards, but the founder’s direct investment in a China-linked AI project served as a top-tier institution’s super check: “A lot of people will think, if Manus can do it, so can I.” From Q2 onward, founders began asking in large numbers whether they could raise Silicon Valley dollars.
• The more interesting signal was that top dollar institutions had already been using data to monitor China’s potential winners. a16z once proactively approached a Chengdu project reportedly in deep cooperation with OpenAI—US AI valuations had risen so fast that Chinese assets looked “unbelievably cheap.”

4. Cheapness Is the Surface; Discount Is the Substance

• Shijie’s breakdown of US-China pricing logic: OpenAI, Anthropic and xAI command an AGI premium. Capital is willing to fund OpenAI’s $16B in compute losses because it is betting that the company becomes “the electricity company of the digital world”—the pricing logic of a sovereign-grade asset.
• China, unable for now to lead the world in SOTA, entered the “read the financials, look at traction, DAU and retention” phase earlier. Capital “does not pay for the foundation-model dream of Chinese teams; it pays only for profitable super-apps or extreme cost-performance.” That makes Chinese application-layer projects especially attractive: open-source DeepSeek and Qwen brought costs down, Manus demonstrated the power of wrapper-product definition and engineering innovation, and the application window is now open.

5. Reverse CFIUS and OIR: Capital and Talent Decouple

• The OIR (Outbound Investment Rule) restricts US capital from investing in Chinese companies in semiconductors, AI and quantum computing, while expanding the definition of a “Chinese company.” A China headquarters, majority ownership by a Chinese person or company, a controlled subsidiary, or financial-data tests could each trigger the rule. Benchmark’s Manus investment was therefore viewed as potentially prohibited, and Manus received a Treasury inquiry in May.
• The result has been a clear chilling effect. Firms such as Sequoia and GGV, which split off their China operations years ago, do not face the same issue; pure-dollar institutions are both excited and cautious about China-linked projects. A tacit formula is taking shape: top Chinese-background engineers and product teams remain welcome, but only if the project fully strips out its Chinese attributes in legal structure, data storage and market focus.

6. 黄敏达’s Warning: Manus Is Not a Model to Copy

• Lawyer 黄敏达’s logic has three parts. First, “in America’s eyes, Singapore is no longer an absolutely trusted place”; moving there “is not a wholehearted pledge of allegiance to the US, and may instead look ambiguous.” Second, cutting the Chinese team or even changing nationalities does not settle the issue—the US really wants to know whether, when the US and China reach an irreconcilable disagreement, the company will stand completely with the US and remain under US control. Third, designing workarounds around the statute is “treating a headache at the head and a foot injury at the foot”; the law is simply a weapon in an evolving contest.
• A post-publication update underscored the point: Chinese authorities were assessing whether Meta’s acquisition of Manus should be subject to review. 黄律师’s framework proved prescient.
• The human cost of choosing sides is stark. A Chinese scholar regarded as an “absolute authority” in embodied intelligence decided, after struggling with US demands to cut ties with China, to return home and start a company. One investor’s verdict: “The previous generation’s globalized entrepreneurship has come to a complete end.” Nobody now dares put the engineering team in China and exploit the cost advantage as before.

7. HeyGen: A Decisive Demonstration of the Narrowest Path

• 徐卓 founded 视界一线科技, HeyGen’s predecessor, in 2020. Three years later, the product had found product-market fit overseas and ARR had surged from $1M to nearly $100M. Benchmark partner Victor Lazarte personally joined the board—“in Silicon Valley, that means a form of bloodline certification.”
• Its pledge of allegiance came before the US entry. Beginning in late 2023, the company lobbied early shareholders including Sequoia, ZhenFund and Baidu to accept buybacks and give up certain board seats, creating capital separation and keeping the company in OIR-safe territory. The headquarters moved entirely to Los Angeles and the founder relocated to the US, creating physical separation. The setup was remarkably similar to Manus, but HeyGen followed the “more thorough and safer” path 黄律师 recommends.
• The head of a dual-currency fund described the three steps: first, users and the market are already overseas and the company has found PMF; second, the growth curve attracts large checks from domestic US-dollar funds; third, the company naturally becomes American. “The problem is that most companies get stuck at step one.”

8. Leaving Silicon Valley Is Capital-Driven: The $300M Ceiling and Expensive US Dollars

• The domestic reality, according to the head of a dual-currency fund: “It is lively before $300M, and you cannot raise after $300M.” The market-based capital pool contains only VCs and speculative money; PE is absent. Any company raising more than $100M faces pressure to cut ties with China—move its headquarters, remove Chinese shareholders, eliminate the Chinese team and control China operating expenses. In principle, the founder must also obtain US status as soon as possible.
• A Silicon Valley-based investor offered a reality check: founders who arrive believing “Silicon Valley has stupid people and too much money” mostly leave disappointed. Zhou Hang, co-founder of LAMA Venture, observes that accent is irrelevant—“a founder can speak with a heavy accent, but cannot express himself poorly.” Many Chinese standouts become less impressive and less conspicuous when placed in Silicon Valley’s talent pool; many arrive, start work immediately and race the clock, but are simply busy with the wrong things. It takes at least 2-3 years to truly adapt to the US.
• Zhou nonetheless praised the new generation, citing the founder of miHoYo, who gave up a large domestic achievement to start over in the US. That reflects an ambition, perspective and sense of purpose unique to this generation of Chinese founders.

9. China’s Brief Primary-Market Spring: Embodied Intelligence Absorbs State Capital

• The surge in sentiment came from several directions: spillover from leading AI stocks in the US provided comps; the reopening of IPOs in Hong Kong and on the STAR Market gave institutions hope of exits and lifted the primary market; Unitree’s Lunar New Year Gala breakout had a long tail; and the IPO of Insta360, along with high-PE performances from Huami and Anker, ignited FOMO in overseas hardware. Across the year, capital converged on four areas: embodied intelligence, AI applications and Agents, multimodality and AI hardware.
• The capital boom in embodied intelligence is visible in the fact that “every recognizable leading company has at least RMB1B in cash on its balance sheet.” The $300M curse does not apply to them because leading companies absorb large pools of state capital and RMB funds. State capital favors long-cycle infrastructure and hardware—chips and embodied intelligence—because “the cycle is long, so the goal is to make as few mistakes as possible,” while it remains extremely cautious on Agents and AI applications because they are “easy to falsify and have high failure rates.”
• Some practitioners counter that “a robot brain is fundamentally similar to an Agent, with a vague wedge and heavy cash burn.” Silicon Valley research also points to a cognitive gap: China owns the hardware supply chain, while the leading exploration at the brain layer remains with US startups and tech giants.

10. AI Applications and Agents: Good Data Draws Big Checks, Middle Eastern Capital Rotates In

• Top dollar funds were noticeably more active in the first half. Sequoia and Hillhouse were reportedly able to invest in 20 projects a month, with $1-2M per deal, although corporate venture arms remained the main force overall. Middle Eastern capital was also increasing: US pressure in 2024 restricted sovereign wealth funds from directly investing in Chinese foundation models, but “that has nothing to do with this wave of AI applications and Agents.”
• The clearest new example is Liblib, founded by 1990s-born ByteDance product manager 陈冕. Daily revenue reached $150K and was still growing; its next-round valuation could reach an astonishing $800M. The gossip is that Alibaba passed on the project at a $500M valuation, then “came back with a second strike” at $800M—another confirmation that genuinely impressive growth means a company “does not struggle to raise money, and does not struggle to raise big money.”

11. Multimodality: China Needs Its Own Nano Banana and Awaits Its Cursor Moment

• Once Google had Nano Banana, “China needed one too.” Kuaishou received a market re-rating on the strength of Kling; among startups, ShengShu Technology is more research-oriented while AIsphere has shown solid commercialization, but the market is still watching for new entrants.
• The most closely watched is Vivix AI, founded by 刘宇 after leaving SenseTime. 刘宇 came out of MMLab at the Chinese University of Hong Kong, founded by 汤晓鸥 and known as the “Whampoa Military Academy of computer vision,” and once commanded several thousand GPU cards. That hands-on experience with low-level compute scheduling and large-parameter models is rare among founders. Vivix’s valuation jumped from $100M-$200M to $1.34B within a year—“the project had never even formally appeared.”
• The key question is timing. Multimodality overall trails language models by 1-2 years and should experience every scaling-law breakout point that language models did. “When will multimodality have its Cursor moment? Who will build multimodality’s DeepSeek?” The market is desperate to identify the answer.

12. Year One for AI Hardware: The Glasses War and the Foundation-Model Entry Point

• Shijie’s periodization: “From 2023 to 2024, the primary market was buying brains—models. In 2025, everyone started buying bodies—the hardware carriers.” The maturation of reasoning, coding and Agents means the brain is ready to connect to hardware, turning the AI assistant discussed for years into reality. The fastest-growing AI hardware globally, Plaud, appears to be the fastest way to connect a foundation model to recording sensors; in substance, it reflects fine-grained engineering around foundation models and the belief that “the next generation of devices must be proactive.”
• Big tech has realized that close-to-body consumer hardware is the gateway to offline data collection and next-generation interaction. Behind the hardware breakout is “a battle for the foundation-model entry point”; wearables that are always on 24 hours a day may be better hosts for foundation models than phones.
• AI glasses became the biggest dark horse. In the spring, Rokid founder 朱明旻 (Misa) put Rokid Glasses on a national leader and onto authoritative media, after which Alibaba, ByteDance, Baidu, Xiaomi and even Li Auto entered the field. IDC forecasts 121% year-over-year growth in China’s smart-glasses market in 2025. Meta’s all-in bet is to rebuild everything itself, from chips and operating systems to AI adaptation and applications, because glasses sit in the face’s core triangle. As one practitioner put it: “Only the camera on AI glasses captures a person’s true first-person view”; necklaces and the pin-style AI Pin cannot reproduce it.
• Policy support matters. In the year closing the 14th Five-Year Plan and connecting to the 15th, the AI Plus initiative explicitly calls for next-generation smart terminals to exceed 70% penetration by 2027. A leading glasses company’s latest financing round reached a remarkable RMB2B; in Q2 2025, AI hardware attracted more funding than software, and total financing for embodied intelligence plus hardware exceeded RMB14.5B.

13. Incubation, Scarcity Marketing and the 27-Year-Old Fund

• VCs are “one step away from moving their desks outside the gates of big tech.” The primary market is trying to persuade technical and product talent from DJI, Płab竹, Anker, Insta360 and Xiaomi to leave and start companies. The popular incubation model is for an institution to provide a clear direction, test execution over 2-3 months, invest if the trial works and walk away if it does not.
• “Invest in young people” has become a consensus. Compared with highly ranked big-tech executives, grassroots operators have stronger equity incentives and “better valuation cost-performance.” 云启 launched a fund this year dedicated to founders under 27: “27 is a magical number.” The founders of Google, Facebook and Microsoft, as well as Bill Gates, were all under 27 when they started.
• FA professionals describe the standard way to inflate a valuation: “Use scarce allocation and rolling valuations to create project momentum, then have a CVC take over at the $100M-$200M stage.” Shijie’s conclusion: a reasonable bubble is a byproduct of industry prosperity and a natural phenomenon, but there will inevitably be cases where investors place the wrong bet.

14. 陈宇’s Window Theory and Three Founder Archetypes

• 陈宇 of 云启资本, who lived through the full autonomous-driving cycle and invested in MiniMax this round, believes the financing window for top projects lasts only a few months. Autonomous driving was the same in 2016: the companies still at the table today all appeared within 1-2 quarters. He believes the final players at the AI table “have already appeared collectively.” Consensus forms so quickly because deep learning had accumulated six or seven years of groundwork before ChatGPT—everything was ready except capital.
• Top Chinese AI founders fall into three archetypes, analogous to Chinese students in the US. The washing-dishes model bootstraps from an existing cash-generating business: Huanfang before its breakout, 蔡浩宇 using game profits to build even its own compute, and Plaud launched with the founder’s own money. The scholarship model earns funding from top global institutions through exceptional performance, as Manus did—“there are very few such targets, and once one is found, the valuation is staggering.” The fundraising model is backed by the national team or RMB funds and deeply tied to national strategy. None of the three lacks money; most of the market’s capital is captured by a small number of projects.
• Two investment styles follow. In hot sectors, investors “pick the tallest dwarf” and prioritize participation; the subtle signal is that a new generation of investors can enter only through the application wave, caring more about the number of portfolio examples on their résumés than actual returns. The other style keeps looking for disruptive projects and is willing to stay out for years. One investor in the latter camp said: “Most projects are simply rebuilding traditional applications from the previous era with AI, like reading newspapers on an iPad, but Toutiao has not appeared yet.” “There are very few institutions and people in China’s market capable of making independent judgments.”

15. Silicon Valley’s Leaderboard: GPT-5 Hits a Wall, Gemini Turns It Around

• The SOTA baton passed rapidly: ChatGPT at the start of the year, Anthropic by year-end, and Gemini in the second half. One secondary-market analyst asked: “Things are changing too fast … the secondary market cannot see the direction, and no one knows how to price next year.” Claude Sonnet 3.7 in May sparked a wave of coding and Agent projects and marked a strategic split: Anthropic abandoned consumer and multimodal efforts to focus on Coding and Agentic because “it is a choice dictated by the competitive landscape—it is a startup, not a giant.”
• OpenAI was training models while pushing productization at full speed. Compute financials circulating in the primary market suggested that a larger share of its compute still went to training rather than inference. After GPT-5, “the ceiling of the scaling law under the existing Transformer architecture may already have appeared.” More than one industry insider has pointed to the toothpaste-squeezing effect since GPT-4. The predictions by former OpenAI scientists Ken and Joe in Episode 48—that “the scaling law is nearing its ceiling and the industry must return to research”—were validated by multiple technical leaders at year-end.
• Google’s comeback came through Gemini 2.5 Pro, which swept the rankings and became the first model ever to top the text, vision and Web-development leaderboards simultaneously. Nano Banana broke out in late August. By Gemini 3 Pro, many practitioners believed its answers had surpassed GPT-5.1 and its multimodal performance was far more stable than GPT’s. A DeepMind researcher described watching the Transformer he helped invent pushed to its limits by OpenAI, which became Silicon Valley’s center of power: “The past two years were extremely painful for Google.” The price was putting other frontier directions on hold and focusing the entire company on LLMs.
• The researcher nonetheless refused to declare victory: “The current victory is only a stage result. We cannot know whether this focus is a complete victory or whether we sacrificed a larger opportunity in other frontier directions.” Google’s current performance is creating real pressure for OpenAI. Sam Altman, unwilling to be tied to Azure, brought in Jensen Huang for Stargate, which observers saw as “OpenAI’s de-Microsoftization”—the most memorable scene in Silicon Valley’s AI war this year.

16. The Stargate Triangle: Oracle, the Junior Partner, Bleeds First

• 姚欣, founder of PPIO, frames Oracle as a “super infrastructure contractor brought in with its own money,” completing the triangle. Nvidia invests up to $100B in OpenAI in tranches, on the condition that OpenAI deploy Nvidia systems; OpenAI uses that money to buy cloud services from Oracle; Oracle then hands those revenues, and possibly more, back to Nvidia to buy hundreds of thousands of GB200s and build data centers. Oracle carries the largest stake in the three-way gamble.
• The market briefly believed that becoming OpenAI’s infrastructure provider meant becoming the next Nvidia. Oracle rose nearly 40% in a single session in September, and Larry Ellison briefly overtook Musk as the world’s richest person. Less than one quarter later, year-end 2025 results showed free cash flow turning negative, while the stock had fallen nearly 40% from its high in Q4—“back to square one overnight.” The more important signal is that independent cloud startups “currently generate less cloud-service revenue than the cost of renting the cards.”
• Nvidia simultaneously spent $20B to acquire the LPU inference-chip assets of Groq, which claims to run foundation-model inference faster and more cheaply than GPUs and is targeting a $100B inference market. Nvidia “used a superpower to neutralize a technical threat and turn a rival into an ecosystem partner.” The narrative shifted from one-way technical challenge to an ecosystem war and alliance-building. At the same time, differentiation among SOTA models is shrinking and lead windows are closing fast: Anthropic’s coding lead was substantial in the first half, then was narrowly matched by Grok, Gemini and other top models in the second half.

17. Meta Buys Momentum; Apple Is Content to Follow

• Many Silicon Valley practitioners and observers believe that after Llama 3, Meta “almost completely collapsed” on the direct foundation-model battlefield. It is behind now and will remain behind for at least another six months—“that is the momentum of a generational gap.” Zuckerberg’s 10x compensation packages to poach talent are viewed as strong medicine, but may damage Meta’s internal consensus culture. 姚欣 learned that four AI teams are running in parallel: the retained Llama team; a 40-50-person team rebuilding closed-source pretraining; the frontier-research team led by Yann LeCun; and a newly created AI-infrastructure team. After LeCun’s departure, the frontier-research group will remain but under new leadership.
• Multiple interviewees reached the same conclusion: “The problem facing every big tech company today, in China and the US, is not innovation, but organization.” Meta is going through the same integration pain Google faced when it merged Brain and DeepMind in 2023-2024. It needs a leader with both technical authority and organizational force; Zuckerberg chose Alexander Wang, but “whether he can shoulder the responsibility is unknown.” Short sellers have also returned to Meta.
• Shijie’s personal view of Meta’s multibillion-dollar acquisition of Manus is that this is not “a stupid-money deal.” Meta may be shifting its center of gravity toward Agents: Manus did not train a model, yet achieved SOTA-level Agent performance. “Could it be that even without building the strongest model, building the strongest Agent execution layer could still write an entirely different script?” Meta is buying momentum and market confidence, much as Nvidia bought Groq’s LPU assets.
• Views on Apple are split. Some believe it has left the table in the AI era, but the DeepMind researcher and Nathan, founder of Plaud, both argue that its talent reserves and depth of product thinking mean that “if it does nothing, so be it—but when it does move, it will carry weight.” Shijie’s analogy is Tencent in China: “not the most aggressive player, but always the one the market expects most, because it can afford to follow rather than lead.”

18. China’s Foundation Models: Falling Behind, Then Overtaking Through Open Source

• Nathan Benaich and Air Street Capital’s State of AI Report 2025 elevated China from a peripheral pursuer to a parallel competitor for the first time: “China is no longer the follower; it is setting the pace in open-source AI and commercial deployment.” The constraints remain objective: overseas frontier models already use Blackwell and will move to Rubin next year, while China can only use gray-market cards. Infrastructure limits the speed of development.
• In open source, DeepSeek, Kimi and Qwen have broadly overtaken the Llama series on multiple key metrics. Zhipu is the commercial-overtaking case: it claims coding performance on par with Claude and ranks third globally, while pricing at one-seventh of Anthropic’s rate and winning sovereign foundation-model orders from numerous Belt and Road countries, partly because of its shareholder structure. In the second half, Silicon Valley startups began switching to Qwen, DeepSeek and K2 across the board—China’s foundation-model sector underwent a qualitative shift in 2025.

19. Three Types of Scaling: Size, Thinking Depth and the Nascent Agent Network

• One industry view is that overseas model companies have shifted primarily to post-training, while China remains focused on pretraining, with Alibaba and DeepSeek investing heavily in post-training. But there is still no evidence that post-training is the future; everyone is searching for the next scaling law.
• Two camps formed around the ceiling at year-end. The wall-builders, represented by LeCun and Ilya, argue that the marginal returns on piling up pretraining compute and data are declining and that the paradigm must change. The non-believers-in-death, represented by Hinton and Dario, argue that the scaling law remains valid but is changing form, with substantial room for inference-side scaling. Claude 3.7 embodies that view, “but the view is controversial.”
• Shijie considers the most valuable argument to be that scaling is moving from model-size scaling to scaling thinking depth, and that the more important future step will be “scaling the network effects created by collaboration among Agents.” The third form is only beginning to emerge. The era of stacking pretraining data and parameters is nearing its end, “the low-hanging fruit will soon be gone,” and the cost of winning that race may be prohibitive.
• Another major technical trend is AI’s move from the digital world into the physical world: language models perform next-token prediction, while world models perform next-state prediction. A Google researcher put it this way: “Today’s LLM world model learns from all the text in the world. Language itself is the boundary of human knowledge, but it is not the boundary of knowledge in the universe … If you cannot interact directly with the universe to obtain a signal, it is hard for your knowledge to exceed the boundary of human language. AGI can approach but cannot arrive.”

20. Back to Commercial Reality: Every Dollar Earned Requires $1.6 of Investment

• Q4 figures circulating in the primary market show OpenAI’s 2025 annualized revenue at $20B, far ahead of the field, but card-rental costs alone at $16B and operating losses at $11.92B—“for every dollar earned, $1.6 has to be invested.” Google and Meta, with their own compute and mature infrastructure, have much better-looking profits; “this is a warning to startups.” Optimists argue that token cost-performance is improving at 10x speed and that many use cases can be fully deployed in 2026, but inference costs are also surging: “training is a drop of water; inference is an entire pond.”
• Control over compute has become a decisive card. The US is securitizing compute assets at scale through bonds and private debt. China cannot buy imported compute, both because it is unavailable and because purchases are prohibited; attention is shifting from “who has more cards” to who can actually access compute. Ant and miHoYo are both building their own. In cloud plus foundation models, the contest is now between Alibaba’s Qwen and ByteDance’s Doubao: Qwen uses open source as a technical wedge to bring application companies onto Alibaba Cloud, creating a “free model, paid compute” model, while compute credits still make up a significant share of many startup investment packages.
• Chinese foundation-model startups are under broad pressure. Discounting compute and inflating PR metrics are widespread; nominal funding and actual cash received diverge, while a market with high valuations is difficult to digest. “The next sharp market chill will be challenging.” The pragmatic route is productization that integrates models and applications, plus the new industry dimension 陈宇 calls model orchestration: maintain an independent identity rather than binding to any major ecosystem, and coordinate multiple models for the best answer. Genspark founder 景鲲 put it plainly: “Everyone can rest assured that our mixture of models will definitely be better than what everyone selects for themselves.”
• Benchmark chasing is no longer meaningful; “growth and retention remain the golden test for foundation-model products.” AI applications will pass through three stages: model-led, model-application separation—the current stage—and integration with industries. The US and China have already diverged: US giants are entering the game-theory board of ecosystem warfare earlier, while China is committed to open source and applications. What is worth watching in 2026 is the emergence of a super-app and a scaling law at the Agent layer. “When it happens, it may radically change our lives.”