Compute, networks, standards: China's technological power revisited
Much of the European debate about Chinese technological power measures the wrong things. Committees joined, proposals tabled, patents declared, households covered: these count mobilisation and reach, not capability and influence. Across a policy brief and two book chapters published this year, and two further papers forthcoming, I argue that the decisive constraints usually sit somewhere other than where the alarm is loudest — and that this is as true of Europe's own position as of China's.
Read the CHERN policy brief on the US-China AI race here, my chapter on the digital China challenge for the EU in The Making of a Global Digital Rulebook here (open access), and my chapter on technical standardisation in Conditioning China's Influence here.
Compute: the constraint has moved to memory
DeepSeek's R1 was read in early 2025 as proof that American export controls had failed. The more interesting story is where the bottlenecks actually lie.
China is not a marginal chip producer. It leads the world in mature logic and is a substantial producer of commodity memory. But it is a trailing producer in precisely the segments that matter most for frontier AI, and capacity cannot simply be redirected, because advanced logic comes from a small number of highly specialised plants. Below 5 nanometres, China would need extreme ultraviolet lithography it does not yet have. A prototype assembled in Shenzhen arrived earlier than many expected; the path to high-volume manufacturing remains roughly consistent with the decade-long estimates.
This is why DeepSeek turned to inference rather than pre-training. The two rely on different hardware: pre-training is compute-intensive, while the binding constraint for inference is memory and bandwidth. That shift weakens the immediate effect of export controls, and Chinese accelerators are competitive for inference. It does not remove the constraint, it relocates it. The binding limit on China's AI accelerator output is now high-bandwidth memory: domestic foundry capacity could supply dies for well over a million Ascend chips a year, while domestic memory supply currently supports a fraction of that.
Two developments have inverted the familiar picture. Access to compute is no longer rationed in Washington alone — Beijing has begun restricting its own firms' access to American accelerators to force adoption of the domestic stack, accepting worse near-term performance for long-term self-reliance. And export controls have become bargaining chips entangled with trade negotiations, which makes them less predictable in both directions.
Rather than attempt a forecast, the brief sets out five indicators worth tracking: the consistency and enforcement of export controls, the relative weight of inference versus training, Chinese progress in high-bandwidth memory, the financial effects of export controls on non-Chinese suppliers, and whether China's AI is simply good enough for the applications that matter.
Networks: Europe closed the coverage gap and missed the architecture gap
The same pattern recurs in mobile infrastructure, where Europe's headline numbers have improved while the thing that matters has not.
Unlike earlier generations, the transformative promise of 5G is industrial rather than consumer-facing. Unlocking it requires three things together: a dedicated 5G core, mid-band spectrum, and private networks tailored to specific industrial sites. The core is the decisive one, and it is invisible in the metrics Europe tracks. An operator can deploy 5G on the most capable spectrum and still run it on a 4G core — which is what the European Commission finds is happening across the Union.
The numbers tell a consistent story, and it gets worse the closer one looks at actual use. Standalone base stations account for 20.9 % of EU mobile base stations, against 34.8 % in China and 36.2 % in the United States. Standalone networks reached 63 % of the European population at the end of 2025, against 93 % in China. And on measured usage, only 2.8 % of European 5G speed tests took place on standalone networks in the final quarter of 2025, against a global average of 17.6 %. Base-station share describes what has been built, population coverage what is reachable, speed tests what is actually used. Europe's position deteriorates at every step.
This matters because the value of these networks lands overwhelmingly outside the sector that builds them. Mobile technologies contributed some USD 1.5 trillion to the Chinese economy in 2025, of which the larger part came from productivity improvements in other industries rather than from the mobile ecosystem itself. There is a reason China reached this conclusion first: Chinese unit labour costs have risen sharply and now exceed those of several Southeast Asian competitors. Industrial connectivity is the substitution — a way of holding productivity advantage as cheap labour disappears.
One detail in China's 15th Five-Year Plan captures the shift precisely. The semiconductor self-sufficiency target Beijing had previously set and missed has been dropped. In its place stands a deployment metric: digital economy value added reaching 12.5 % of GDP by 2030. China has moved from measuring success by what it manufactures to measuring it by how deeply digital infrastructure penetrates the economy.
Europe's one genuine success here is rarely recognised as such. Germany reserved the 3.7–3.8 GHz band for local assignment in 2019 and priced it on a published, deliberately low formula; by November 2025 it had issued 484 assignments across manufacturing, logistics and research. It is the single instrument most directly responsible for Europe's position in private networks — and it has never been generalised across the Single Market.
Standards: three kinds of power, and what they attach to
Technical standards look like engineering documents, which is why their political weight has been so easy to miss. They are voluntary specifications, written by engineers, binding on nobody. Yet they operate as market power, deciding whose patents earn licensing income and whose products need redesigning; as security policy, because lock-in in critical infrastructure creates decades-long supplier dependencies; and as normative infrastructure, because assumptions about privacy, surveillance and human oversight are quietly written into systems everyone uses.
China's influence here is substantial and uneven. It participates in more ISO technical committees than any other country while holding fewer secretariats than Germany; in the IEC it ranks sixth for secretariats. Its clear lead is in the ITU's standardisation sector — the one forum where governments rather than private bodies carry the weight, which is itself revealing. Beyond the formal bodies, China spreads de facto standards through translation programmes, more than a hundred standardisation and mutual-recognition agreements, and infrastructure projects where the standard travels with the loan.
Two findings cut against the prevailing alarm. First, China's system excels at mobilisation but converts it into implemented standards far less reliably; Chinese assessments themselves point to definitions too vague to test against, overlapping committee mandates and voluntary standards firms have little incentive to adopt. Second, a standard on its own does very little. Its economic consequences appear when it is attached to something else: a test method, a certificate, a procurement rule, a subsidy condition, a loan. Private actors can make some of those attachments. States can make many more, and deliberately. That, rather than what happens in the committee room, is where China's advantage lies.
Europe's more urgent problem is its own delivery. The Commission's evaluation puts the average development of a harmonised European standard at close to six years, against a Chinese target of eighteen months. The average age of harmonised standards cited in the Official Journal is twelve years. European standardisation organisations answered one of eight standardisation priorities in 2022 and three of nine in 2023.
Two open files
The temptation is to copy the Chinese coupling of standards to industrial policy. That would be the most expensive mistake available. Europe's system works because its independence from politics is what makes its standards trusted, and trusted far beyond Europe's own market. A Europe that traded that away would end up slower than the largely private American system, less coordinated than a state-led one, and no longer the reference for anyone.
What Europe should do instead follows from what it controls, and two legislative files are open right now.
The revision of the Standardisation Regulation due in 2026 should set binding deadlines for the steps the Commission itself controls, turn standardisation requests from invitations into negotiated commitments with named deliverables, and keep common specifications as an emergency instrument rather than the standing answer to a delivery problem. Alongside it, Europe should pay for participation — making standardisation work a reimbursable cost and building a succession pipeline before an ageing expert generation retires.
The Digital Networks Act, proposed in January 2026, harmonises spectrum for public networks but does not establish a framework for local, non-public assignments of the kind Germany operates. That gap should be closed while the file is open. Europe should also measure and incentivise network architecture rather than population coverage, use public procurement in seaports, hospitals, transport hubs and public research sites to create the demand operators cannot underwrite, and settle the security framework for private industrial networks — which the 5G Toolbox was never designed to cover — before the Act is adopted.
The underlying argument
All of this rests on a position I have argued for some years. Technological decoupling from China is neither feasible nor desired, and the interdependence is genuine in both directions. If self-reliance is not the goal, Europe can only avoid being constrained by its dependencies if it does two things at once: reduce strategic dependencies where it can, and remain technologically indispensable to China where it cannot. The second half requires a shift in perspective — away from cataloguing European weaknesses and towards identifying and preserving European strengths. The virtualisation of networks is a warning on exactly this point: it moves value away from the radio hardware in which Europe is strong and towards the cloud and AI layers in which it is not.
Three forthcoming policy briefs develop these arguments further: China's technical standardisation power in future technologies for the Friedrich Naumann Foundation, China's quiet industrial policy for the Friedrich Ebert Foundation, and Industrial 5G and the new digital divide for the China in Europe Research Network.