Weekend Food For Thought WFFT
On This Week's Menu: The US Economy, Private Equity and Private Credit, Electricity 2026 Analysis and Forecast to 2030, Humanoids Are Not Just Cars With Legs, Dunnin-Kruger Effect and more...
Hello from Lisbon,
I hope you had an interesting and productive week.
C. S. Holling stated that: “Resilience is the ability of a system to absorb change and still persist.”
May you find insights and knowledge from the inputs below that helps you absorb change and still persist…
1 Getting Visual
2 If You Read One Thing Today - Make Sure it is This
3 Consequential Thinking about Consequential Matters
4 Big Ideas
5 Big thinking
6 Surround yourself with builders
1 Getting Visual
The Long View: Global Goods Exports belongs to the process innovators - The share of global goods exports that isn’t from Asia or Western Europe is small…
The Long View: Novel innovation has been very profitable as “Software ate the world” and powered US Big Tech to record profits. ”For big companies, the regular cost of doing business has been in decline for a while. For the median S&P 500 company, SG&A as a share of revenue has been in gradual decline since Dotcom, while COGS as a share of revenue has dropped even more dramatically, going from ~67% to about 55% (and falling) over the last 30 years. The decline in operating expenses has (unsurprisingly) coincided with a steady rise in operating margins, as well, to now historical highs. Now, a lot of this is obviously explained by the prevalence of high margin software and tech businesses increasingly dominating the index, but since we’re talking about the median large cap, that’s not the whole story. Another part of the story is tech itself making companies more efficient, as they’ve been able to scale their organizations, while keeping operating costs in check. The reward for good work, though, is more work. And having pushed margins to historical highs, the market may be asking “what have you done for me lately,” especially as AI disruption increasingly enters the zeitgeist. The question now, then, is whether companies continue to cut costs (or at least keep them in check), or do they now begin to invest towards the promise of a better tomorrow, perhaps even by increasing their tech budgets (to the delight of beleaguered SaaSCos and AI upstarts alike).” - A16z
Signal: Industrial policy gone wrong? Cutting power inputs while ramping up usage…The US saw an uptick in installed power plants under the IRA, it has plateaued since 2024 just as data center builds ramped up…
Key Input: 𝗧𝗵𝗲 𝗮𝘃𝗲𝗿𝗮𝗴𝗲 US city electricity 𝗽𝗿𝗶𝗰𝗲 𝗽𝗲𝗿 𝗸𝗪𝗵 𝗵𝗮𝘀 𝗶𝗻𝗰𝗿𝗲𝗮𝘀𝗲𝗱 𝟯𝟬% 𝘀𝗶𝗻𝗰𝗲 𝟮𝟬𝟮𝟮. From about 14.5 cents per kWh in 2022 to roughly 19 cents per kWh today. The line on this chart is not just noisy volatility. It is a structural step up.
Key Question: Do you have the Minerals? Behind every AI model and cloud server sits a complex web of minerals that make modern computing possible. From semiconductors to cooling systems, these materials form the backbone of digital infrastructure. Semiconductors are the “brains” of AI data centers—and they are highly import dependent. The U.S. is 100% reliant on imports for arsenic, fluorspar, gallium, germanium, indium, and tantalum used in chip production. It also imports 85% of its platinum and 36% of its palladium needs, both critical for chip manufacturing. While silicon, the base material for chips, has less than 50% import reliance, many of the trace elements that enable advanced computing are entirely foreign-controlled. Beyond chips, server boards and circuitry require a range of conductive and precious metals. The U.S. imports 64% of its silver and 73% of its tin, both vital for soldering and electrical conductivity. Copper—essential for wiring and connectivity—has a 45% import reliance. Tantalum, used in capacitors, is 100% imported. Beyond chips, server boards and circuitry require a range of conductive and precious metals. The U.S. imports 64% of its silver and 73% of its tin, both vital for soldering and electrical conductivity. Copper—essential for wiring and connectivity—has a 45% import reliance. Tantalum, used in capacitors, is 100% imported. Meanwhile, data storage components such as magnets and drives depend on rare earth elements, with 80% import reliance. Barite—used in storage-related applications—has also more than 75% reliance. China’s Commanding Share Currently, China dominates the production of most of the critical minerals used in data centers. This near-monopoly has become a major concern for other nations, with the U.S. government currently pushing for increased domestic production of these materials. In addition to being the leading producer, China also controls much of the refining capacity for many of these minerals. For example, around 90% of rare earths are refined in China. In the race to dominate AI, access to critical minerals may prove just as important as technological leadership.
Learn more: https://www.visualcapitalist.com/the-critical-minerals-powering-the-ai-boom/?mc_cid=828e30ab98&mc_eid=814179b7ef
Macro Spotlight 1: The US Economy - change is not an event, it is a process and you can observe the elements that makes up a shifting process. US Y-on-Y Job Growth has been hinting at an economic slowdown (or a major productivity leap) while the labor demand/cost in key industries like Healthcare and Education remains sticky and along with tariffs and trade/import disruption could lead to Stagflation. Will the recent GDP drop be an outlier or the first of several big steps down?
Macro Spotlight 2: The US Economy - change is not an event, it is a process and you can observe the elements that makes up a shifting process. A look at startup space tells a similar story - “Layoffs, AI, and the future of startup work: Data below shows layoffs across the ~60,000 or so US startups using Carta. Let’s run through some recent history: 2018-2019 was chill. April 2020 was decidedly not chill. Massive layoff spike as COVID anxiety ran rampant. Late 2020 into 2021, things calmed down again. Actually better than that - layoffs were low and new offers were plentiful. Interest rates rose for the first time in forever, March 2022. Uh oh. Biggest-ever spike in layoffs across startups in January 2023. This was coming to terms with the new reality across startups (high interest rates, low funding, not fun). Also ChatGPT launched in Nov 2022, just noting. Since the peak, layoffs have trended lower over 2024 and 2025. Not quickly, no sharp decline. But the trend is steadily lower. 𝗪𝗵𝗮𝘁 𝗛𝗮𝗽𝗽𝗲𝗻𝘀 𝗡𝗼𝘄? A lot of people have predicted that AI will result in large waves of startup layoffs. I think this is a little vague. AI itself has not pushed layoffs upwards (clearly), although it may be playing a role on the margins. The real change is in hiring. Startups are building with fewer people than they did 1, 2, or 3 years ago. The median seed-stage startup on Carta now employs 5 people, down from 10 or so in 2021. At Series A, the same drop was 26 --> 17. Fewer hires --> harder job market --> more people staying in current jobs --> fewer open roles --> harder job market, and so on. Others accept this reality but say that’s okay as long as there is a Cambrian explosion of small companies employing 2-10 folks. Enough of those startups and boom, employment overall holds steady or grows...I’m skeptical. One fact: probably harder to get hired at a startup now than it has been in a decade or so.” - Carta
Demographic Spotlight: A Country for Old Men (for now) - More than 70% of all the wealth in the US is held by people over 55. Wealth Transition is just getting underway and the related 2nd and 3rd order effects are going to be massive - Are your family, your business and your portfolio prepared?
Learn more: https://www.federalreserve.gov/releases/z1/dataviz/dfa/distribute/chart/?utm_
2 If You Read One Thing Today - Make Sure it is This
For actual observers of the “House of PE” and it’s related cousin “Private Credit” the signs of trouble have been there for a while - as I ponder the state of affairs the “Blue Owl” is howling with plenty of rapidly changing explanations for it’s apparent troubles (go check out the excellent “Accredited Insight” Substack for the latest on this from the hardworking Layla Kunimoto:
Here Pitchbook - who, as to their name, generally are firmly on the sales side of PE and VC - takes a look at “Private Equity’s Exposure to the Software Reckoning” and while they acknowledge some issues they generally take the glass half full of Cool-aid view - go check it out in full (half empty) - plenty of potential for 2nd and 3rd order effects that may rear its head in unexpected places and ways - both positive and negative - from this reset, always good to explore both sides…do it here:
https://pitchbook.brightspotcdn.com/b0/1f/01e7659644729883076873196a72/q1-2026-pitchbook-analyst-note-private-equitys-exposure-to-the-software-reckoning.pdf
Some Takeaways
“Private equity exposure to software is elevated: Software represented around 18%
of US PE deal value in 2025, increasing the asset class’s sensitivity to the current public valuation reset.
Public software multiples have compressed: Multiples are down more than one standard deviation from their eight-year average, driven by AI disruption.
Structural frictions limit the pace of wholesale AI-driven displacement: Enterprise security, compliance, switching costs, and user retraining suggest evolution rather than immediate obsolescence for many SaaS businesses.
Periods of relative underperformance in technology-focused PE have historically created attractive deployment windows: This dynamic is particularly relevant when valuation compression outpaces fundamental deterioration.
Software has become one of the core sectors within US buyouts.
Over the past decade, it has represented approximately 14% of total PE deal value and 11% of total PE deal count. In 2025, that exposure increased meaningfully, with software accounting for roughly 18% of total deal value, while its share of deal count stayed flat.
In other words, at the very moment that public valuations are resetting, private equity’s capital concentration in software is elevated. Conversations with industry participants make clear that both AI-driven opportunity—through greater efficiency and enhanced product value—and AI-related disruption risk have been central to underwriting discussions over the past two years. What feels different now is the perceived pace and magnitude of change. Technological progress has appeared faster than many anticipated, and the disruption risk seems more material than previously expected.
That uncertainty has contributed to a noticeably more cautious tone among investors.”
“Yet focusing only on disruption risk overlooks the structural strengths embedded in many software businesses. A software company is more than its code.”
“With many public software names trading near five-year lows, markets appear to be pricing in an aggressive disruption scenario. For private equity investors, this environment may represent opportunity rather than existential threat. Firms willing to underwrite durable franchises and integrate AI into existing platforms may drive a new phase of growth.”
Public software valuation reset
For much of the past decade, software was viewed as the ideal private equity asset.
Recurring revenue, pricing power, approximately 70% gross margins, and the potential for 30% operating margins at scale created a compelling financial profile. The capital-light model and ability to support leverage made software particularly attractive in a low-interest-rate environment.
Software valuations surged in 2020 and 2021 amid aggressive monetary stimulus and rising inflation. Many investors came to view the sector as structurally resilient, even as a potential inflation hedge. Empirically, valuations displayed a clear inverse relationship with US five-year Treasury rates. As rates declined, multiples expanded; when rates rose, valuations compressed. The framework appeared stable, and many investments made in 2024 and 2025 likely incorporated the expectation that future rate cuts would support renewed multiple expansion.
That relationship has since broken down. Despite rates moving modestly lower, public software multiples recently fell to more than one standard deviation (SD) below their eight-year average. AI disruption risk is clearly the primary driver of the recent reset. The open question is how much of a discount is appropriate to reflect AI-driven disruption risk.
Amid this debate, it is important to recognize the psychological dimension. Investors in 2024 and 2025 were still anchoring expectations to the extraordinary conditions of 2020 and 2021.
Anchoring is a well-documented cognitive bias: When exposed to extreme reference points, people tend to adjust too slowly as conditions normalize.
Peak-era valuations became an implicit benchmark, making recent multiples feel like an attractive buying opportunity with the prospect of further rate cuts in the years ahead.
Net net, the risks in IT-focused private equity reflect uncertainty around AI and its implications for the software-as-a-service (SaaS) model itself. Investors are questioning whether traditional application layers can be commoditized, whether AI- native platforms will reshape established workflows, and whether pricing power will erode as functionality migrates into foundation models. For the first time in more than a decade, the central debate is not about growth rates, but about structural durability.
History, however, suggests caution before declaring disruption fatal. Markets often overestimate the near-term impact of transformative technologies. Autonomous driving was once expected to cripple incumbent automakers and ride-hailing platforms. Instead, both industries continue to operate, and many incumbents are investing in the same technology that investors once presumed would displace them.
Technological shifts rarely eliminate established players overnight. More often, they reshape competition and reward adaptation.
In software, two barriers deserve attention: enterprise security and user retraining. Large enterprises, particularly in regulated industries, can take quarters or even a year to approve new software. Security reviews, data governance, and integration testing create institutional friction.
AI may increase that friction. Leaders must evaluate not only functionality but also model risk, data exposure, and embedded vulnerabilities. In environments where a single breach carries material financial and reputational cost, caution is rational.
There is also a behavioral constraint. AI tools are improving rapidly, but AI-generated output often reflects the statistical center of its training data. It can accelerate workflows, yet it does not replace leading-edge expertise. Users tend to remain with the tools they know, especially when productivity depends on familiarity.
Consider Microsoft Office. Free alternatives with near feature parity, such as OpenOffice, have existed for years. Yet enterprises remain anchored to the incumbent ecosystem: file formats, integrations, macros, workflows, and training. Switching costs are not merely financial. They introduce operational risk that few organizations accept lightly. The practical test is simple: Attempt to move your organization to OpenOffice.
The idea that AI will instantly displace entire categories of software assumes away these frictions. Change will likely occur through augmentation and gradual integration rather than wholesale replacement. For now, compliance requirements, organizational inertia, and human habit remain stabilizing forces.”
IT-focused PE performance
Technology-focused private equity firms delivered exceptional performance from 2010 to 2020, outperforming diversified peers by more than 700 basis points over this period. That era was driven by multiple expansion, durable growth, and scalable software economics.
The cycle has turned. After the sharp valuation reset in 2022 and a brief rebound in 2023, rolling one-year IRRs show technology-focused PE firms lagging diversified PE firms again in 2025. Sentiment and narrative risk around AI have weighed on marks.
Historically, periods of relative underperformance have created compelling deployment windows, particularly when driven more by valuation compression than by fundamental collapse. If current fears around AI disruption prove overstated, today’s environment could represent one of the more attractive entry points in years. The differentiator will be strategy. Not all software businesses are equally at risk of disruption, and not all managers are equally positioned to underwrite an AI-integrated future.”
What gets software out of its historic valuation reset?
Markets are volatile. Consensus builds, crowding follows, and reversals overshoot.
Today, sentiment has driven public software valuations to multiyear lows. Dislocations of this magnitude rarely persist indefinitely. Over time, fundamentals reassert themselves. Durable franchises with real cash flow and embedded customer relationships begin to look less like falling knives and more like mispriced assets.
Not every company will recover. Some business models will prove structurally impaired. But those that demonstrate durability will likely catalyze the turn through buybacks, take-privates, and strategic consolidation.
Severe resets are uncomfortable in real time. In hindsight, they are often remembered as the moments when long-term investors looked past the volatility and found rare opportunities.”
3 Consequential Thinking about Consequential Matters
The IEA has dropped their ‘Electricity 2026 - Analysis and forecast to 2030’ report. It covers global key trends and developments in this consequential area - it’s 225 pages long but worth at least a skim in full - go do it here:
https://iea.blob.core.windows.net/assets/b73798cb-e452-42b9-9d8a-07542de7a041/Electricity_2026.pdf
Some Takeaways
“Global power demand growth continues to rise rapidly as the Age of Electricity gathers pace, supported by the increasing electrification of industry, transportation, and the buildings sectors. Growing consumption is also coming from some of the most dynamic segments of global economies, such as artificial intelligence (AI), data centres, and evolving technological innovations.”
Global electricity demand is forecast to increase at a brisk average annual rate of 3.6% over the 2026-2030 forecast period, supported by rising consumption from industry, electric vehicles, air conditioning and data centres. Worldwide electricity demand grew by 3% year-on-year in 2025. This followed growth of 4.4% in 2024, when intense heat waves and strong industrial activity in many regions boosted electricity use. Looking ahead, annual demand growth over the next five years is set to be 50% higher on average compared with the average across the previous decade.
For the first time in three decades, excluding periods of crisis‑related disruption, global electricity demand outpaced economic growth in 2024 in what is set to become a broader trend in the coming years. Despite a slight reversal in 2025 due to weather conditions that affected electricity demand, a fundamental shift in the longstanding relationship between electricity demand and economic activity is set to be a defining feature of the forecast period. Through 2030, electricity consumption is projected to grow at least 2.5 times as fast as overall energy demand.
Emerging economies continue to be the main pillar of demand growth, accounting for nearly 80% of additional electricity consumption through 2030.
While India and Southeast Asia are increasingly set to drive rising energy demand over the coming decade, China is forecast to remain the single largest contributor to global electricity demand growth through 2030, accounting for close to 50% of the increase.
Over the next five years, China alone is expected to add demand equivalent to the total electricity consumption of the European Union (EU) today, with average growth of 4.9% annually. This is close to its 2025 pace of 5 but slower than its 6.5% average over the past decade.
India and Southeast Asia’s share of electricity demand growth among emerging economies is forecast to rise substantially by 2030, driven by robust economic growth and rapidly rising demand for air conditioning, which is set to boost both annual consumption and peak loads.
Electricity demand growth in advanced economies is accelerating again after 15 years of stagnation.
This resurgence signals a new era in which electricity is a major energy input to some of the most dynamic drivers of global economies, such as artificial intelligence (AI), data centres and advanced manufacturing. In 2025, advanced economies accounted for almost 20% of global electricity demand growth, up from 17% in 2024. We expect this share to remain near the 20% level on average over the forecast period, driven by expanding industrial activity and the continued growth of data centres, electric vehicles and other end‑uses of electricity. In the United States, electricity demand rose by 2.1% in 2025 and is projected to grow by nearly 2% annually through 2030, with around half of the total increase driven by the rapid expansion of data centres. After rising by less than 1% in 2025, electricity demand in the European Union is expected to grow more strongly. Assuming a moderate rebound in industrial activity, EU demand is forecast to increase by around 2% per year through 2030 – although consumption is not expected to return to 2021 levels before 2028. Many other advanced economies – such as Australia, Canada, Japan and Korea – are also expected to see faster electricity demand growth through 2030.”
Half of the world’s electricity is forecast to come from renewables and nuclear by 2030.
Total generation from renewables is overtaking coal, in line with previous IEA forecasts. With solar PV generating record amounts of electricity, renewable output rose rapidly in 2025, virtually matching the level of coal‑fired generation based on the latest available data.
This was despite weaker hydropower output in some regions and lower‑than‑average wind speeds, particularly in Europe, which tempered overall growth in renewable generation. Renewable output is forecast to grow by about 1 000 terawatt-hours (TWh) annually through 2030, with solar PV alone accounting for over 600 TWh. In percentage terms, renewable generation is forecast to rise at an annual rate of 8% per year. Renewables and nuclear are together expected to account for around half of global electricity generation by 2030.
Nuclear generation set a new record in 2025 and is set to continue rising steadily through 2030.”
The Age of Electricity requires a fast and efficient expansion of grids and system flexibility to securely and cost‑effectively integrate a changing mix of generation, demand and storage. Variable output from solar PV and wind continues to expand quickly, with their share of global generation set to rise from 17% today to 27% by 2030. Meanwhile, newer sources of demand – such as electric vehicles, heat pumps and highly concentrated loads, such as data centres – are expected to grow rapidly.
At the same time, more than 2 500 gigawatts (GW) worth of projects – encompassing renewables, storage, and projects with large loads, such as data centres − remain stalled in grid connection queues worldwide. Since grid investment has lagged well behind investment in generation capacity, many power systems are already experiencing rising congestion‑related curtailment.
Meeting forecasted electricity demand through 2030 would require annual grid investment to increase by roughly 50% by 2030 from today’s USD 400 billion, alongside a significant scaling up of grid-related supply chains.
At the same time, grids built for peak capacity often have substantial unused capacity during off‑peak periods. As grids and flexibility rise up the policy agenda, making more efficient use of existing systems can help relieve congestion and accelerate integration while grid expansion efforts continue.”
“Utility-scale battery deployment is accelerating rapidly, becoming a significant source of short-term flexibility. While conventional power plants remain the primary source of power system flexibility, the growing fleet of large‑scale batteries is playing a rising role in supporting security of supply. The strong growth is especially notable in regions with rapidly rising shares of solar PV and wind in electricity generation. Markets such as California, Germany, South Australia, Texas and the United Kingdom have all seen strong growth in utility‑scale battery capacity in recent years. Battery costs continue to fall, enhancing their competitiveness, but efforts to reduce market barriers and address integration challenges can help unlock their full potential.”
China’s industrial demand growth is increasingly driven by electrified manufacturing outside heavy industry…
Heavy industry, which covers iron and steel, non-metallic minerals and the chemicals and petrochemicals industries, has seen its contribution to total electricity demand growth decrease in recent years. While its share in demand growth was between 15-20% in the 2010s, this has fallen to well below 10% in the last five years, and is expected to decline further to around 6% over the forecast period, despite support measures and significant growth in the chemical and petrochemical subsector through 2030.
Non-heavy industries have followed the opposite trend, reaching a more than 40% share of China’s demand growth between 2020 and 2025, from 33% in 2010-2015. This growth has been driven by subsectors such as machinery, non-ferrous metals, transport equipment, textile and food industry. New energy products also had a very significant impact on this growth, as noted in our Electricity 2025 report.”
“Renewable electricity generation grew in India by 20% in 2025, posting their absolute strongest annual increase (+82 TWh) on record. Growth was led by solar PV generation (+24% y-o-y), with 33 TWh of additional generation. This is a notable rise from the already strong 15% growth seen in 2024. Wind power generation was up 28%, while hydropower rose by 14% amid improved hydrological conditions. By contrast, nuclear generation decreased by 1.6%. Given strong increases in low-emissions sources in a year when the country’s overall electricity demand growth rate remained relatively muted at 1.4%, coal- fired generation fell by 3.2%, after a 5% rise in 2024. Gas-fired generation similarly is estimated to have declined by 9%, after rising by 6% in 2024.”
“In the United States, electricity demand remained strong, rising by 2.1%. Electricity generation from solar PV rose by 70 TWh (+26%) in 2025. Total renewable output increased by more than 8%, with higher solar PV generation alone accounting for about 80% of the gains. By contrast, wind generation growth slowed to 2.9%, compared to 7.6% in 2024. Coal-fired electricity was up by a sharp 13% y-o-y in 2025, rebounding strongly following a contraction of almost 3% in 2024, while gas-fired generation declined by 3.6%.
In the European Union, solar PV and wind generation combined surpassed fossil-fired generation in 2025, marking a milestone. Renewables share in total electricity generation approached 48%. Solar PV generation rose by a substantial 22%, overtaking hydropower to become the second-largest source of renewable electricity, behind only wind energy. At the same time, amid reduced wind speeds, wind generation was down 2.6% y-o-y and hydropower fell by around 13% due to reduced rainfall. Nuclear generation remained stable in 2025 amid robust output in France and a number of other countries. Even though EU electricity demand growth was less than 1%, lower wind and hydropower generation contributed to higher gas burn in the power sector, which increased by around 8%. Equally, weather effects led to a more moderate 6% decline in coal‑fired generation, compared with an average annual decline of more than 20% in 2023 and 2024.”
“Strong solar PV growth through 2030 remains a common trend across the regions…
In many parts of the world, electricity generation from ever-cheaper solar PV is showing the largest growth among supply sources. Over our forecast period, the share of solar PV in the total electricity generation mix will surpass 10% in many major economies, and even reach 20% in some of these countries. In terms of absolute and relative growth in solar PV, China leads the way among major economies. The solar PV share surpassed the 10% threshold in China in 2025 and is set to exceed the 20% mark in 2030. After a year-on-year increase of 360 TWh in 2025, solar PV generation is forecast to rise by 320-360 TWh every year out to 2030, meeting around 60% of the average annual demand growth. We expect all of China’s additional electricity demand in 2026-2030 to be met by low-emissions sources, renewables and nuclear energy.
The total increase in solar PV generation in China over the next five years will be larger than the combined increase in solar PV generation in the rest of the world.”
4 Big Ideas
Two explorations with some overlap - first Poe Zhao of the excellent ‘Hello China Tech’ Substack explores why “Humanoids are not just cars with legs” but how Chinese carmakers cite a 60% supply-chain overlap and 70% software reuse. Plenty of Big Ideas explored in this piece go check it out here:
Secondly, China Talk’s Nick Sorvino asks: Is China Cooking Waymo? And he goes on to explore how Ii terms of international expansion, Chinese firms are way ahead of the American competition. Chinese companies have worked out Autonomous Vehicle (AV) deployment deals with more than thirteen countries. The US: two. Chinese companies are also exporting something closer to a full autonomy stack — vehicles bundled with cloud services, AI traffic management systems, and road sensors. Read it in full here:
“Humanoids are not just cars with legs” - Some Takeaways…
“Fifteen Chinese car companies have entered the humanoid robot field, according to Kaiyuan Securities, a Chinese brokerage. They are building, investing, and partnering their way into a technology they believe runs parallel to what they already do. Two numbers underpin this belief. CITIC Securities, one of China’s largest investment banks, estimates that the hardware supply chains for smart vehicles and humanoid robots overlap by more than 60 percent: sensors, motors, batteries, cameras, and compute chips serve both product categories. Separately, XPeng claims that its robot and car reuse 70 percent of the same AI software stack, spanning perception algorithms, motion planning, and domain controllers.”
“Together, these figures form the structural foundation of the car-to-robot pivot. If the overlap is real and sufficient, automakers carry a meaningful advantage over pure-play robot startups. If the remaining gaps, in dexterous manipulation, dynamic balance, and close-range human interaction, prove to be where the true difficulty concentrates, these companies are reorganizing their engineering cores around a premise that physics may not support.
The question extends beyond China. Tesla’s Optimus program rests on the same logic. Elon Musk has halted production of the Model S and Model X to free factory lines for robots and tied his compensation milestones to delivering one million Optimus units. But China’s electric vehicle market, with dozens of brands locked in a punishing price war, is generating a wider and faster set of experiments.”
What the Bridge Promises
XPeng’s approach is the most direct. On February 3, two days after IRON fell, the company merged its autonomous driving and smart cockpit divisions into a single unit called the General Intelligence Center. One foundation model will handle perception, decision-making, and motion control across the car, the cockpit, and the robot, drawing on shared sensor hardware, domain controllers, and AI infrastructure. The reasoning follows a clean line: if a car can perceive its surroundings, plan a route, and execute, a robot needs the same capabilities in a different physical form.
Li Auto, the EV maker that built its brand on range-extended family SUVs, is moving faster and risking more. On January 26, CEO Li Xiang held a company-wide meeting and announced that Li Auto would build humanoid robots. The autonomous driving department was dissolved. Its chief, Lang Xianpeng, the architect of Li Auto’s self-driving strategy, was reassigned to lead the new robot team. AI evaluation and data labeling groups followed him.
The timing reveals the motivation. Li Auto had paused robot development in late 2024, calling the technology premature. What shifted in the fourteen months since was not any breakthrough in robotics. Li Auto posted a net loss of 625 million yuan (roughly $86 million) in Q3 2025, ending eleven consecutive profitable quarters. Revenue fell 36 percent year-over-year. When reports of the company’s robot ambitions began circulating in late January, its U.S.-listed shares rose roughly 7 percent in a single session.
NIO, China’s premium EV brand, rejected the builder’s path entirely. CEO Li Bin has indicated that NIO will not develop robot hardware, at least for now. His stated focus is on NIO’s proprietary chip program and on positioning the company’s technology as a supply chain asset for the broader robotics industry. Rather than assembling an internal robot team, NIO’s venture capital arm has backed at least three embodied intelligence companies over the past year: LimX Dynamics, a Shenzhen humanoid robot company that raised $200 million in its latest round;Linghou Robotics, a component and assembly firm that has shipped over 2,000 units; and Yuanli Lingji, a venture backed by Ant Group, the fintech company behind Alipay.
Seres followed a playbook that already transformed its business. The company manufactures Huawei’s best-selling smart vehicles: Huawei provides the software and brand, Seres provides the factory. That partnership lifted Seres from near-irrelevance to China’s second most valuable automaker, a roughly twelvefold increase in market capitalization over five years. When Seres entered robotics, it replicated the formula. In October 2025, it signed a preliminary cooperation agreement with Volcengine, the cloud and AI arm of ByteDance, TikTok’s parent company, to co-develop intelligent robot technology. Under the proposed arrangement, ByteDance would contribute AI algorithms, computing power, and multimodal models, while Seres would contribute manufacturing capability and industrial test scenarios. The deal remains a framework agreement. Specific projects, financial commitments, and timelines have not been disclosed.
Four companies. One shared foundation: the technology stack behind a smart car will carry over to a humanoid robot.
Where the Bridge Ends
Evidence from three directions suggests the bridge is shorter than its builders assume.
The first is the raw technical distance between driving on roads and moving through rooms. Traffic is constrained by lanes, maps, and norms. Rooms are not. Lanes are painted. Traffic signals follow fixed patterns. The physics of forward motion at speed is well-modeled. Humanoid robots face the inverse: unstructured spaces where they must open doors, grip small objects, climb stairs, and navigate around obstacles that shift without warning. The dexterity, force control, and whole-body coordination required for these tasks sit well beyond anything in a vehicle’s control architecture. Musk has argued that Optimus is harder than major vehicle programs like the Model X, though he’s also said Starship is harder.
The difficulty surfaces even in the step before robotics. Both XPeng and Li Auto reorganized partly to achieve what the industry calls cockpit-driving fusion: combining the car’s interactive AI with its autonomous driving system into one unified intelligence. This alone is proving stubborn. The two systems operate under fundamentally different constraints. A smart cockpit tolerates errors; a wrong song recommendation is an inconvenience. An autonomous driving system demands millisecond-level determinism; a single miscalculation is a safety event. Merging the two onto shared computing hardware risks one starving the other of processing power at a critical moment. Additional obstacles compound the challenge: different safety certification standards, different software update cadences, and different engineering cultures between cockpit and driving teams, as detailed in reporting by Jiemian, a Chinese financial publication. If unifying AI inside a car poses this many challenges, coordinating a walking, grasping, stair-climbing body represents a different order of complexity.
The second signal comes from talent. If the car-to-robot bridge were solid, the best engineers would stay inside automakers and walk across it. Many are heading the other direction. The founder of XPeng’s original robotics subsidiary departed to start his own company. Former autonomous driving leaders from XPeng, Li Auto, Huawei’s automotive division, and Horizon Robotics, a major Chinese autonomous driving chip company, have co-founded at least five embodied intelligence startups. These engineers recognized the robot opportunity and concluded they would move faster outside a car company.
The third signal is competitive. Robot companies that never relied on automotive technology are running ahead. Unitree, whose robots danced on the Spring Festival Gala, China’s most-watched annual broadcast, reported shipping 5,500 units in 2025. Agibot, a well-funded humanoid robot startup, has reported revenue approaching one billion yuan. LimX Dynamics has released a modular robot platform and what it describes as the industry’s first embodied agentic operating system, enabling robots to plan and execute multi-step tasks without human direction. These companies were built for robotics from the ground up. They are delivering product while car companies are still redrawing org charts.
The Cost of Crossing
The automakers that committed to building robots internally face a bind that tightens with each step forward.
XPeng and Li Auto are pulling their strongest autonomous driving talent into robot programs. Li Auto’s position is the most exposed. Its top self-driving executive now leads robots. The team that built the company’s competitive edge in driving intelligence has been distributed across three new groups. If the shared technology premise holds, this reallocation will prove farsighted. If the premise falls short, Li Auto risks weakening its standing in the car market, where Huawei-powered competitors plan to field 25 new models in 2026, without closing the gap with robot startups that hold multi-year leads in hardware and deployment.
NIO’s strategy inverts the risk profile. Funding the robot supply chain rather than building robots avoids cannibalizing car engineering. But if embodied intelligence becomes the next platform shift, an investor captures a smaller share of value than a builder.
Seres is replicating a partnership model that delivered extraordinary results in vehicles. The structural question persists: when the core intelligence belongs to a partner, margins tend to follow the intelligence provider, not the hardware assembler.
These are Chinese variations of a question approaching every global automaker. Musk has committed Tesla’s next chapter to the same transfer thesis. China’s fifteen parallel experiments will produce data faster and in greater volume than any single company could generate. The early readings counsel caution. The overlap between cars and robots is real in sensors, compute hardware, and perception software. It thins rapidly once a robot must manipulate objects with precision, maintain balance on uneven ground, or work safely within arm’s reach of a person.
He Xiaopeng said IRON’s stumble looked like a child learning to walk. A closer parallel may be a highway racer entering an obstacle course. The vehicle shares some relevant capabilities. Certain skills carry over. But the terrain is entirely different. And terrain, ultimately, determines who finishes.”
“Is China Cooking Waymo?” - Some Takeaways…
“At times, this piece reads like a typical “US vs China” article, but in fact we’re seeing more of a “co-opetition” dynamic Kevin Xu highlighted in the AI industry. In fact, the perhaps more interesting aspect is how the line between “Chinese AV company” and “US AV company” blurs in practice. Chinese AVs use NVIDIA chips, Waymo uses Chinese-made Zeekr vehicles, and Uber and Lyft partner with Chinese AV firms internationally, not to mention critical minerals. The industries are too tangled for neat distinctions… but let’s try to untangle them anyway.”
“The US has Waymo. China has its “big three”: Baidu’s Apollo Go, WeRide, and Pony.ai (whose founders ChinaTalk interviewed last year). There are other potential players in the US, like Amazon’s Zoox and Tesla (RIP GM’s Cruise). But right now, Waymo is the only American company operating a scaled, paid Level-4 robotaxi service, which enables vehicles to handle all driving tasks within specific operating zones. China also has BYD and Xiaomi with L2 driving features (who could transition to L4 soon), and many more robovan, robus, robodelivery, and robotruck companies on track for L4 deployment.
In aggregate terms, China appears to have the edge in overall deployment. An analysis by SCSP suggests Chinese autonomous-vehicle operators have collectively logged roughly 149 million autonomous miles, compared to around 106 million miles for US firms — a roughly 1.4 to 1 advantage.
But mileage comparisons are limited. Companies report different levels of autonomy, mix supervised and driverless miles, and disclose data unevenly across jurisdictions. Ridership is a different way to look at it, where China has completed ~30 million rides, versus ~20 million for the US.”
Different Services
Ridership itself misses a big part of the story, because China’s AV industry extends beyond passenger ride-hailing. By the end of 2024, more than 6,000 driverless delivery vehicles were reportedly operating across 100+ city zones. Companies such as Neolix, Zelos, Meituan, JD Logistics, and Alibaba’s Cainiao are actively piloting or scaling operations for shipping, food delivery, and street-cleaning vehicles.
The US, by contrast, doesn’t yet have road-going (as opposed to sidewalk-going) driverless AVs deployed, with companies like Nuro still limited to pilots and R&D fleets. They do have an estimated 3,000 to 5,000 sidewalk AVs, but China has many of these also, and they are a much simpler technology.
China also seems to be leading in autonomous trucking.
The International AV Market
The divergence between US and Chinese approaches to autonomous vehicles becomes even more pronounced in their international expansion strategies.
US Companies (i.e., just Waymo currently) have deals with:
Chinese companies have deals with:
The international lead for Chinese companies is larger than it appears, because countries already integrated into China’s Digital Silk Road infrastructure are natural targets for the next wave of expansion. Egypt, for example, is attempting to ditch Cairo by building the New Capitol, with extensive Chinese infrastructure support for “smart city features,” the kind of project that sounds ripe for AV deployment. Similar rumblings have been true for Oman. WeRide also struck a partnership with Grab, the biggest ride-hailing app in Southeast Asia, which positions it well for expansion into the entire subcontinent.”
Middle East
China has the clearest advantage in the Middle East.
What makes the Middle East such a natural fit is that many governments there can reconfigure the built environment and the regulatory environment in parallel to accommodate AVs. The US strategy, thus far, seems to be to drop AVs into existing roads and regulatory systems and hope for the best.
For example, Saudi Arabia signed a deal with WeRide in 2024 to deploy robotaxis across the Neom megaproject. With help from the Chinese, the planned city is being built with dedicated infrastructure for autonomous vehicles. The UAE is doing similar things. WeRide has reportedly accumulated 1 million km of operational mileage in Abu Dhabi already.”
Who has leverage in the AV supply chain?
China’s Leverage
LiDAR uses laser pulses to create detailed 3D maps of a vehicle’s surroundings, measuring distances to objects. It’s the spinning sensor thingy typically mounted on top of autonomous vehicles. China controls ~90% of the global LiDAR market. Firms like Hesai, RoboSense, Huawei, and Seyond account for the bulk of automotive-grade LiDAR shipments.
Waymo, importantly, produces its own LiDAR in-house, so China can’t cut it off. But most of the AV industry lacks that level of vertical integration, and because Waymo’s LiDAR is proprietary, new companies entering the market will likely be beholden to Chinese suppliers. (An AI analogy is Google’s TPUs for AI models; they can produce these for themselves, but a new entrant into the AI game would likely be reliant on NVIDIA’s GPUs.)
Batteries are another potential choke point. Most robotaxis are electric, which pulls AVs directly into the EV battery supply chain that China dominates end-to-end. CATL and BYD together account for over half of global EV battery installations, with companies like Tesla, BMW, Ford, Volkswagen, and Toyota all using them. Even Waymo’s next-generation robotaxis use Chinese-made Zeekr vehicles powered by batteries from CATL.
Interestingly, CATL was designated as a “Chinese military company” by the Biden-era DoD and added to the blacklist for government or military usage. This blacklisting doesn’t apply to commercial vehicles…yet.
These advantages translate directly into cost. Public estimates routinely put Waymo’s current robotaxis at roughly $130k-$150k per vehicle, once sensors, compute, and the base car are included. It is purported that they spend $40-50k just on sensors (like LiDAR), since they are paying a premium to produce it themselves. Chinese robotaxi platforms, by contrast, are cited at $30k-$50k all-in, due to state subsidies and the preexisting car manufacturing prowess.
US Leverage
If China were to withhold access to LiDAR or batteries, how could the US respond?
Most directly through NVIDIA chips, but not the standard Hopper and Blackwell series GPUs used in AI data centers. Unlike data center GPUs designed for training, AVs use automotive-grade systems-on-chip (SoCs) like NVIDIA DRIVE Orin and Thor — integrated platforms that combine CPU, GPU, and dedicated neural network accelerators optimized for real-time inference, safety certification, and lower power consumption. Baidu uses dual Nvidia Orin X chips, Pony.ai uses four Orin chips, and WeRide recently deployed Nvidia’s Thor platform.
However, the technical barriers differ significantly between AI and AV chips. For AI GPUs, cutting-edge nodes (2-5nm with EUV lithography) are essential to achieve the compute density required for training workloads, but autonomous driving chips face lower node requirements. Current AV chips like Nvidia’s Orin operate on relatively mature ~8nm processes, since AV workloads prioritize deterministic latency, power efficiency, safety certification, and software integration overjust raw compute density. This means Chinese domestic foundries like SMIC could theoretically produce decent AV chips on 7-12nm nodes without accessing advanced lithography equipment, whereas comparable AI training chips would require the cutting-edge processes that remain out of reach.”
Conclusion
China’s AV sector is performing strongly — controlling key parts of the supply chain, dominating international expansion, and scaling up deployment domestically.
But here’s a puzzle: why did both Pony.ai and WeRide experience brutal post-IPO crashes? Pony.ai fell from its debut price of $15 to $4.18 before recovering to around $16-17, while WeRide plummeted from $15.50 to around $9-10. Their valuations of approximately $7 billion and $3 billion pale in comparison to Waymo’s estimated $110+ billion valuation.
I suspect these conditions can coexist. China’s unique political economy has proven it can (1) dominate a global sector without (2) guaranteeing it is financially lucrative. A similar dynamic has played out with EVs. China produces 70% of global EV output, yet most Chinese EV makers operate at a loss. State-backed capacity buildout created severe overcapacity and price wars, with automaker margins falling from 5.0% in 2023 to 4.4% in 2024 despite surging volumes.
The brutally competitive domestic market may explain why these AV companies are racing overseas. If you strike a deal with the UAE and are the only AV company able to operate there, there is some room to breathe away from the vicious competition within China.”
5 Big thinking
We live in a time where it is important to adopt the Dunning-Kruger lens as you look out into the morass of ‘noise’, most originating from the “Peak of Mount Stupid”…The Dunning-Kruger effect is a type of cognitive bias where smarter people tend to underestimate themselves, while ignorant people are more likely to think they are brilliant and let everyone know about it. Keep that in mind as you scan for who to listen to…
6 Surround yourself with builders
Have a Great weekend when You get to that stage,
Sune