Summary

AI data centers are competing aggressively for high-priced HBM, pulling consumer electronics into a prolonged supply squeeze. Apple has absorbed the pressure for roughly 3 quarters, while 李楠 says 库克 once lobbied the U.S. government for access to key CXMT capacity—evidence that “Apple’s cushion is not as deep as people think.” Supply-side estimates put even the most optimistic case at 2 years; pessimistic scenarios have cited 1 year, 4 years, and 5 years, while data-center demand is already being projected out to 2030.
The smartphone market could shrink 10%–15%, but the real story is K-shaped divergence rather than the average. 庄明浩 estimates that brands defending their position may decline by just 2–3 points, while weaker players could fall 20%–30%; 李楠 believes neither RMB1,000 phones nor the ultra-premium segment will disappear—the ones truly dying are those stuck in the middle. Xiaomi’s upmarket push was prescient but could still fail.
Low-priced hardware can no longer compete through simple spec cuts; the next battleground is joint efficiency gains across the OS, network and cloud. 李楠 envisions squeezing the entire stack in reverse, from the Linux kernel and memory compression to low-cost data plans and cloud storage, with the goal of “making 4 GB of memory run as smoothly as someone else’s 12 GB.” Huawei is relatively more proactive, while Xiaomi and OV are more constrained because they do not control the core OS; Huawei itself remains limited by lower layers and geopolitics.
AI phones look as if they have been forced into maturity this year, but vendors must still use them to create a reason to upgrade. 庄明浩 sees more original product thinking from Xiaomi, while Huawei may have the most complete implementation, yet no one has truly answered what an AI smartphone is. The 6–9-month innovation gap between large and small vendors will rapidly narrow: “It cannot deliver this year, but it has to.”
Voice interaction will be adopted when recognition accuracy crosses an acceptable threshold. 李楠 estimates that “80% of users are lazy”; once accuracy reaches roughly 97%, users may no longer want to type. Angry Miao plans to launch a keyboard with an integrated microphone array this year, but he also believes reliance on existing voice models is not a long-term moat: “Innovation is actually prediction.”
Hardware companies need to shift from subsidized scale to “healthy scale” and turn their installed base into recurring service revenue. The replacement cycle has stretched from under 12 months to nearly 4 years, making battery replacement, cloud services and Token spending potentially more important than one-off handset sales. Nintendo can pass through costs with weak hardware and strong games, while the roughly RMB7,000 Steam Machine exposes the true price of unsubsidized hardware.
Rising costs have not killed investment in AI hardware; they have increased the value of small teams using market change to break through. Consensus sectors such as dexterous hands, world models and AI toys have become a game of “stacking the blinds first,” but founders from DJI and Insta360 still attract capital. Shenzhen’s supply chain plus AI has enabled a one-person team to crowdfund and deliver a smart ring, validating the idea that “the thing small companies fear least is a market that keeps changing.”

Deep dive

1. HBM Is Dragging Consumer Electronics into a Long-Term Supply Reshuffle

Alan relayed the mainstream industry explanation: AI data centers need vast amounts of HBM, and the higher selling prices of high-bandwidth memory are pulling capacity away from consumer storage products. Memory and drives then rise in price up the supply chain. Being able to buy at list price on Singles’ Day may count as a good outcome.
李楠 is more inclined to think Apple is “out of moves”: he says 库克 once asked the U.S. government to ease restrictions and allow Apple to use key supply-chain capacity from CXMT. Given Apple’s caution in managing political and business relationships, making such a request would mean the company is feeling the pain firsthand.
庄明浩 focuses on the timing. The trend had been clear across the industry since Q3 last year, yet Apple held out for at most 3 quarters before adjusting its strategy ahead of its biggest annual launch event. Price increases are therefore no surprise; the real variable is when each vendor raises prices, cuts specs or changes its product mix.

2. A Smartphone Downturn Could Put Brands on a 2%–30% Spread

庄明浩’s scenario is a market decline of roughly 10%–15%. If the average decline lands at 13%–15%, brands that defend their position may fall by only 2–3 points, while weaker players could head toward 20%–30%. In a zero-sum installed-base market where “your gain is my incremental volume,” the brand gap will widen quickly.
李楠 does not see this downturn as sudden. The smartphone industry’s inflection point was already clear in 2017; nearly 10 years later, the consequences are simply continuing to play out. Xiaomi’s push upmarket was prescient but may still fail. OV is also under pressure, though not necessarily more than value-focused Xiaomi.
His K-shaped view is that neither the low end nor the ultra-premium segment will disappear; “the ones truly dying are those stuck in the middle.” Premium brands have gross-margin buffers, while RMB1,000 phones serve a degree of social function. But the market will also produce awkward devices that roll back their specs, reuse last year’s configuration and simply add the words “new version.”
李楠 was unsparing toward established vendors: “Double-digit declines today are deserved.” In his view, even if the Doubao phone sells only modestly, it at least offers something new. With Apple’s foldable phone already “99.9% predictable,” the industry does not need another spec refresh; it needs effective innovation.

3. The Low-End Phone’s Way Out Is to Squeeze Efficiency from Software, Networks and the OS

K-shaped consumption is supporting low-end demand, while supply-chain inflation is raising the minimum hardware cost. That contradiction can no longer be solved through spec cuts. 李楠 believes AI, Web Coding and already widespread high-speed networks will inevitably be used to squeeze more efficiency from software.
The Redmi K90 made the pressure visible. The standard version was attacked by users over its price on launch night; at 9 a.m. the next morning, 雷军 posted on Weibo offering a RMB300 coupon and, “perhaps for the first time,” a heavyweight publicly said that memory and drives had risen too much. A price increase may save an individual product, but it cannot solve the low-end market’s problem.
李楠’s proposed path includes Linux kernel optimization, memory compression, low-cost data plans and expanded cloud storage, ultimately delivering “the smoothness of someone else’s 12 GB with 4 GB of memory.” In the past, when software ran slowly, users could add memory. For at least the next 3 years, the industry will have to “squeeze the entire software and AI stack in reverse.”
The program noted that phone vendors began adding memory compression to new OS releases last year. Huawei is seen as relatively more proactive, with stronger end-to-end control, but it remains constrained by Qualcomm, Android’s lower layers and geopolitics. Even if Xiaomi implements similar optimizations, its control may be only a discounted version of Huawei’s. 庄明浩 puts his confidence in vendors’ technical claims at 40%.

4. AI Phones Are Being Forced into Maturity as the Timing Gap between Big Vendors and Startups Closes

庄明浩 believes AI smartphones were already a trend but are now being “forced into maturity” by cost pressure. Apple has also registered AI.apple.com. Xiaomi has relatively more original product thinking, while Huawei may have the most complete implementation, but no company has yet reached the level of imagination such a device should have.
AI capabilities must take on the job of stimulating upgrades this year: “It cannot deliver this year, but it has to.” Phone vendors serve tens of millions or even hundreds of millions of users and used to move 6–9 months slower than smaller companies because of compliance, legal and organizational burdens. That window is now set to shrink sharply.
庄明浩 says Google I/O’s demonstrations of writing apps, AI voice-input keyboards and self-built Widgets were all activities that once looked more like startups’ work. Pressure will push major phone vendors to enter directly. Meanwhile, tier-1.5 vendors already have AI product managers born in the 2000s running key businesses. The real constraint is that their organizations do not allow them to pursue overly aggressive form-factor innovation.

5. Voice Will Absorb GUI Functions, while Mice, Keyboards and Film Could Become Emotional Assets

李楠 says Angry Miao is developing a keyboard with an integrated microphone array and plans to launch it this year. The product is intended to eliminate ambient desktop noise so users can talk naturally with AI. But the team increasingly believes that dependence on Doubao’s traffic or existing voice models is not a moat; future open-source multimodal models will also have strong speech recognition.
The host offered a realistic counterexample: speaking alone at home can feel embarrassing, and users may even feel that “my thoughts cannot keep up with my mouth.” 李楠’s product judgment is that 80% of users are sufficiently lazy. Recognition does not need to reach 100%; at roughly 97%, users may stop wanting to type.
He therefore redefines innovation as prediction rather than “an idea that sounds reasonable”: watch the slope, assess growth potential and estimate the ceiling. Voice replacing GUI to a very large degree is, in his view, a curve with a steep slope and a high ceiling. The mouse will remain focused on FPS gaming in the near term and is not readily replaceable.
庄明浩 used vinyl as another possible outcome: after years of decline, it returned to growth, even though many buyers do not own a turntable. Film cameras and mechanical keyboards were also cited as examples of nostalgia and emotional value. 李楠 calls this return a “Renaissance”: “The truly great things among old objects, after being filtered by time, will come back.”

6. Software Differentiation Determines Who Can Pass BOM Costs Through; the Subsidized-Hardware Model Is Coming Due

The team used an AI music guitar to illustrate a mismatch between hardware and software cycles. An internet company took nearly 2 years from project approval to finished hardware, while AI music capabilities were advancing by leaps each quarter. Before the product had even launched, the company had to plan a second generation and include component inflation and even “a war in the Taiwan Strait” in its BOM risk.
庄明浩 sees Nintendo as an unusually comfortable hardware brand in this cycle. It has competed for years with “junk hardware and extremely strong software and gameplay,” leaving it little pressure even when prices rise. 李楠 added that Nintendo has raised prices as well.
The Steam Machine is the counterexample. Its roughly RMB7,000 price drew fierce criticism from players in China and abroad, while Steam explained repeatedly that it could no longer squeeze the price down. The discussion exposed how consoles such as Xbox and PS5 have long subsidized hardware to monetize game sales, leading to a simpler conclusion: in the future, hardware itself will at least need to make money.

7. As Upgrade Cycles Lengthen, Installed-Base Quality Matters More Than Shipment Scale

At the height of the smartphone boom, the replacement cycle was under 12 months; today it is close to 4 years. If vendors cannot drive rapid upgrades, official battery replacement is not merely a repair service—it extends the time users remain inside the brand’s ecosystem. Annual post-sale revenue may be modest, but its gross margin is high.
李楠 frames cloud services as a shift from one-time payment to recurring payment. A user who once paid roughly RMB7,000 for excess local compute could instead pay about RMB1,000 upfront and then consume roughly RMB3,000 in Token usage over time. That resembles installment payments while making the product easier for lower-priced-device users to enter.
庄明浩 uses this to emphasize “healthy scale.” If growth is no longer the primary objective, Xiaomi and other vendors must assess whether their installed base has genuine spending power. Used-device, leasing and refurbished-device data are also growing. Product innovation will increasingly span hardware, systems, services, payment models and usage patterns.

8. Doubao Has Won Chatbot Users but Still Hasn’t Solved China’s Subscription Problem

庄明浩 says Doubao could copy leading overseas models by charging different prices for the best model, usage quotas and scenarios. It could also pursue non-productivity use cases such as voice, voice personas and companionship. The first route is mature but may not fit the domestic market; the second requires a much larger leap in innovation.
Chinese users have historically shown limited willingness to pay for productivity tools, but Doubao’s user scale gives it the capital—and the “latitude and luxury”—to experiment. Its professional version ultimately chose a productivity-oriented pricing path. That route is easier to explain and price, but it must bear the risk of adapting to the market.
The host cited foreign media reports saying that Doubao’s revenue, despite roughly 300M users, is “worlds apart” from ByteDance’s, while explicitly saying it was unclear how much monetization pressure the model team was under. At least from the outcome, the commercialization task facing the product and operations teams is now impossible to avoid.
庄明浩 pushes the question down to organizational design. The model team may be tasked with catching up in Coding, multimodal generation and real-time generative world models, while the Doubao product team carries the commercialization KPI. Agents, Coding tools and Token Plan form a new battleground, but how much overlap they have with users previously educated by Chatbot remains unclear.

9. AI Hardware Funding Continues, but Consensus Sectors Have Become a Blind-Bet Game

庄明浩 describes the current capital market as: “Don’t look at the cards; stack the blinds first.” Dexterous hands, world models and last year’s AI toys could all become consensus sectors at some point. By the time the media is covering them frequently, the chips have already been placed—and even strategic investors may no longer have a seat at the table.
Rising costs affect the decision, but they do not outweigh the positive buff of a founder’s track record. If core members of DJI or Insta360 leave to build small hardware products, investors will still take the meeting. As long as the fundraising target is not excessive, someone will invest first, followed quickly by a second and third investor filling out the round.
Alan mentioned an entrepreneur interviewed by journalist 韦鹏 who used AI, prior hardware experience and Shenzhen’s supply chain to build a smart-ring company essentially alone, successfully crowdfund it on Kickstarter and begin shipping. 庄明浩 then said an extremely small team should be able to build a highly original smartphone. 李楠 said he “wouldn’t build a phone,” having already watched “wave after wave of pioneers go to their deaths.”
庄明浩’s contrarian view is that “small companies fear competition that is too predictable,” because a large company can crush them with capital alone. Small companies fear a market that keeps changing much less. Rising component prices have made the traditional scale-for-price model ineffective, giving smaller companies room to answer the question anew.

10. Supply Will Take at Least 2 Years to Ease; No Clear Downtrend Is Likely before 2030

The program reasoned from the factory cycle. Only a handful of vendors can actually produce scarce components. Even if roughly $1T in funding cited by Korea were invested now—the episode was not certain about the exact figure—building production lines, completing testing and ramping yields would take at least 2 years, while data-center demand would continue to grow.
The customer mix matters even more. The bulk of vendors’ business is already B2B, serving OpenAI, Anthropic, Nvidia, Microsoft, Google and Amazon. The program also cited Micron’s decision to cut its consumer business last year. Even if total output rises, consumer electronics could continue to lose share.
李楠 breaks the excess profits into 3 sources: HBM’s technological lead, the tight oligopoly formed by Samsung, SK Hynix and Micron, and privileges and licenses. His attribution is explicit: long-running restrictions on China’s semiconductor capacity have reinforced the monopoly, and global consumers ultimately pay the cost.
庄明浩 says even the most optimistic case is 2 years; the original pessimistic estimates listed 1 year, 4 years and 5 years. 李楠 believes it is basically unlikely to see the relevant curve move downward before 2030. The hope rests on Chinese suppliers such as CXMT: “If China enters, give it a 70% gross margin and let it earn for 3 years—that absolutely cannot happen.” His direct advice to consumers is to buy now while they can; waiting may not bring a discount and could instead mean paying more.

Building a Company in Stealth | Travis Kalanick with a16z Ant Lingbo's 沈宇军 on Robot-Native Foundation Models and Data Scale-Up