Baidu's Robin Li and Xiaomi's Lei Jun told the APEC CEO Summit on November 17, 2026 that artificial intelligence spending is running ahead of profitable adoption, according to Reuters and Channel News Asia. Li warned the industry could face a real AI bubble if investment fails to translate into business profits; Lei said few companies have closed a commercial AI loop that actually earns money.
The timing is deliberate. Baidu reported another revenue decline the day before, CNA reports, while global markets watched Nvidia earnings and fallout from the Megaspeed export-control scandal. For builders, the speeches are not bearish futurism—they are an on-record admission that tokens burned do not equal revenue booked.
TL;DR: what each CEO argued
| Question | Robin Li (Baidu) | Lei Jun (Xiaomi) |
|---|---|---|
| Core risk? | Capex without profit → bubble | Huge spend, few paying scenes |
| Fix? | High-value applications atop models | Physical AI (phones, cars, robots) |
| Infrastructure idea? | Token Exchange to pool inference supply | Factory robots that work 8-hour shifts |
| Tone | Urgent but still calls AI inevitable | Anxious practitioner, not skeptic |
What Robin Li said at APEC
Tower of blocks beside a scorecard, evoking AI bubble warnings at APEC
Reuters paraphrases Li at the Shenzhen panel:
- Most firms have not seen significant revenue or profit from AI adoption after heavy spending.
- If that gap persists, "the concern that AI might be a bubble could well become a reality."
- He splits the industry into chips, cloud, models, and applications, arguing competition is fiercest at the foundation while value must be captured upstairs.
CNA's livestream summary adds sharper wording: AI penetration into the application layer has fallen short of expectations, and "for most, the answer is no" when asked whether adoption raised profits.
Li also promoted Baidu's Token Exchange, described in Reuters as matching token providers with enterprise and individual inference customers to ease compute shortages. That is a product story for Chinese cloud buyers—and a mirror of why DeepSeek capped free chat one day earlier citing insufficient compute.
TechNode's November 17 recap ties Li's APEC comments to organizational redesign themes from Baidu World, where he argued companies must change workflows, not just buy GPUs—similar to explainx.ai's write-up of agent harness adoption friction.
What Lei Jun emphasized
Sina Finance syndicating 财联社 quotes Lei on the same stage:
- Four years of model progress, but "very few" companies earn a real commercial AI loop.
- AI phones and AI cars have not yet produced a revolutionary benchmark product, though he believes one will emerge.
- Xiaomi's answer is embodied AI: factory robots on 8-hour internships, self-charging twice per shift—adjacent to Unitree and humanoid narratives but anchored in manufacturing revenue.
Lei's frame complements Li's: if cloud and model capex is global, margin may only appear when AI touches atoms—cars, devices, and plants—not when another chat wrapper ships.
Why this matters if you build with models
1. ROI is now a C-suite line, not a blog footnote. When Baidu and Xiaomi leaders—who run large inference stacks—say profits lag spend, procurement teams will ask for unit economics per feature. Price evals, agents, and RAG against revenue or cost saved, not demo quality alone. See Baseten's efficient inference guide for how teams cut $/token without smaller models.
2. Application layer beats parameter races. Li's four-layer stack matches how explainx.ai covered ERNIE 5.5 open weights and Tencent HY4: new checkpoints keep landing, but distribution and workflow decide who gets paid. If your roadmap is "integrate the newest API," expect harder questions after APEC.
3. Compute rationing is the other half of the bubble warning. DeepSeek's November 16 quotas and Li's Token Exchange pitch describe the same physics: demand curves hit HBM and power ceilings before software matures. Budget reserved GPUs, prefix caching, and smaller fine-tunes the way Together AI documented for long-horizon agents.
4. China-US rhetoric is background, not the lesson. Some Hacker News commenters argued Chinese CEOs benefit if US markets fear an AI bubble. Whether or not you accept that motive, the operational claim—enterprise adoption is slow—is consistent with Western CIO surveys and Americans' negative AI sentiment polls we summarized in November.
IMF and market jitters: why CEOs spoke now
CNA places Li's warnings beside IMF Managing Director Kristalina Georgieva telling AFP in October 2026 that a loss of investor confidence in AI returns could create financial stability risks if equity valuations reprice sharply. That macro line matters for developers selling into public companies: CFOs under a falling Baidu revenue line will treat inference burn like opex, not R&D romance.
The same CNA piece ties sentiment to Nvidia's first post-Huang-earnings session and Megaspeed headlines. None of that changes vLLM release notes, but it changes who signs your PO.
If your pitch is "we'll fine-tune later when models improve," expect pushback. The counter-pitch Li and Lei implicitly endorse: ship a workflow that earns margin with today's weights.
Robin Li's "smart organization" checklist (panel details)
Sina Finance's panel transcript relay lists three organizational principles Li repeated from recent Baidu events:
- Break job boundaries because models are generalists—one person can span roles with agent help.
- Digitize tacit knowledge stuck in senior employees' heads.
- Close workflow loops so human–agent steps produce measurable data.
Those are change-management requirements, not API features. explainx.ai's SWE-Cowork benchmark shows the gap: agents look strong in sandboxes, but ticket systems, IAM, and procurement still gate production.
For a 25-person startup, the lesson is smaller but real: document which agent steps require human sign-off, export traces, and tie each automation to a KPI leadership already tracks—otherwise APEC-style speeches become your buyer's excuse to freeze pilots.
Embodied AI on the same stage
The Shenzhen panel was not only LLM economics. Unitree's Wang Xingxing, UBTECH's Zhou Jian, and CATL's Zeng Yuqun spoke on robots, batteries, and eVTOL power density in the same program summarized by Sina Finance.
Lei Jun's 8-hour factory robot anecdote is the embodied counterpart to Li's application-layer speech: margin shows up when AI moves atoms and when after-sales scales. That rhymes with ScaleVLA's 552B robot backbone post and VicOne robotics CVE research—physical AI is where capex, safety, and revenue meet.
What developers are saying on Hacker News
The Baidu, Xiaomi CEOs thread passed 100 points the same afternoon. Useful practitioner reports—not verified facts:
- origini (comment 50532106) argued adoption is slow because security, connectors, and environments take months, not because models lack quality—the timeline problem, not investment levels.
- SecretDreams (50532691) said tools are strong but ideas are repetitive, with many teams rebuilding the same agents.
- slowin (50532388) described personal build-it-yourself stacks that work for individuals but do not scale to enterprise procurement cycles.
- jrflo (50532228) noted Baidu and Xiaomi are not frontier labs, comparing the warnings to Samsung commenting on OpenAI—skepticism worth weighing when you cite the quotes.
Disagreement is part of the thread: some readers treat the speeches as geopolitical signaling, others as belated honesty after years of vibecoding hype.
How this connects to Baidu's model strategy
explainx.ai already tracks Baidu's open-weight push and claimed benchmark disputes:
- ERNIE 5.5 Flash Base for self-hosted inference.
- WebDev arena rank claimed vs verified when marketing outran third-party tables.
- Unlimited OCR as an application-shaped release Li would probably call "high-value."
Li's APEC speech does not retract those releases. It reframes them: models are table stakes; margin lives in vertical workflows—tax, maps, ads, robots—that survive a bubble scolding.
Policy and summit context
APEC placed AI beside US-China diplomacy. CNA notes Xi–Trump talks scheduled later in the week with AI risk on the agenda. Executive warnings about bubble risk may inform negotiation tone, but they do not change export rules on their own.
Builders exporting to either market should still read Commerce remote-access IFR coverage and Malaysia chip loan pauses as the compliance side of the same capex boom Li described.
What to do this week
| If you… | Then… |
|---|---|
| Sell AI features to enterprises | Attach $ saved / $ earned metrics to pilots; expect procurement to cite APEC-type speeches. |
| Burn tokens on free tiers | Model quota risk like DeepSeek's; move production to metered APIs with caps. |
| Chase newest checkpoints | Ship one workflow that survives without weekly model swaps—pricing, support, eval gates. |
| Compete in China | Watch Baidu Token Exchange docs for domestic liquidity; don't assume US neocloud pricing. |
Hourglass of tokens beside adoption warning, echoing AI bubble risk at APEC
Related reading
- DeepSeek daily usage limits, November 2026
- Moonshot Kimi science pivot as coding benchmarks flatten
- Tencent HY4 Flash and WorkBuddy pricing
- Americans' negative AI sentiment across five polls
- Baseten efficient frontier inference guide
- China US model distillation reporting
Quotes and financial figures reflect Reuters, CNA, TechNode, and Sina Finance coverage accessed November 17, 2026 UTC. Baidu SEC filings and earnings releases override media paraphrases when they conflict.
