Weekly briefing: AI demand moves beyond the processor

The early stages of the AI boom were dominated by processors. This week offered a clearer view of the hardware needed around them.

There were faster optical links, another step in China’s memory production and power components intended for increasingly dense server systems. At the same time, proposed European procurement rules showed how quickly industrial policy can change the commercial position of an otherwise suitable component.

These developments will reach engineering teams through availability, qualification work, thermal limits and sourcing restrictions. They are less visible than a new processor, but often harder to work around once a design has been approved.

Optical links move closer to the processor

ECOC 2026 in Málaga brought a series of demonstrations intended to support larger AI systems.

Marvell showed optical technology built using a 2 nm process. The demonstrations included 400 Gbit/s per lane PAM4, an 800G ZR and ZR+ module with MACsec security and 1.6T coherent designs. It also showed a 102.4 Tbit/s platform using optics packaged close to the switching silicon.

These remain demonstrations rather than results from customer systems. Their direction is still important.

As accelerator clusters spread across more racks and buildings, electrical connections face practical limits in reach, power consumption and heat. Optical design can no longer be left until the main computing architecture has already been decided.

Tower Semiconductor and NewPhotonics supplied a firmer production signal. The companies said they have begun volume shipments of optical chips supporting 800G and 1.6T connections.

The devices combine lasers, amplifiers, modulators and photodetectors on Tower’s silicon photonics platform. NewPhotonics also plans to begin volume shipments of a 6.4T chipset during the first half of 2027.

China adds another source of advanced memory

Chinese memory manufacturer CXMT announced that its fifth generation DRAM platform has entered mass production.

The company also introduced two 24 Gbit LPDDR5X products for smartphones and portable electronics. CXMT claims the new process can produce at least 50 per cent more dies from each wafer than its previous platform.

Those figures come from the manufacturer and will require independent evidence from production yields, reliability testing and customer qualification. The start of mass production is nevertheless more significant than another laboratory result.

An additional supplier could improve availability in a market still dominated by Samsung, SK Hynix and Micron. It also creates a more complicated approval decision.

Commercial availability does not automatically make a memory device a usable second source. Timing, package details, controller support, firmware configuration and environmental performance must still match the design.

Export controls and geographical restrictions add another layer. A component may be technically suitable but unavailable for a particular customer, programme or destination.

Power demand reaches the component catalogue

TDK extended six ranges of snap in aluminium electrolytic capacitors to ratings of up to 500 V DC. The company is targeting power supplies used in AI servers and other demanding switching applications.

It is not as dramatic as a new accelerator or optical platform, but it shows where the pressure from AI infrastructure is moving.

Higher distribution voltages can reduce the current needed to deliver a given amount of power. They also place greater demands on insulation, spacing, protection circuits, conversion stages and component life.

The 500 V rating is only one part of the selection. Ripple current, internal temperature, cooling, endurance and fault conditions will decide whether the capacitor is suitable for the finished supply.

The wider point is that power density cannot be solved at the processor alone. It travels through the complete system, from incoming power and conversion to connectors, circuit boards, cooling and protection.

Production volume increases the cost of poor traceability

Cognex introduced a new reader for identifying wafers and semiconductor panels. The In Sight 1750 combines controlled lighting with recognition software intended to read damaged marks and difficult surfaces.

Identification equipment does not usually receive the attention given to process nodes or new devices. It becomes more important as production volumes and packaging complexity increase.

A missed or incorrect code can separate a wafer or panel from its process history. The result may be a stopped machine, manual checks or uncertainty over which production stages the material has completed.

The important measure will not be the presence of AI in the reader. It will be the read rate and false read rate achieved on the marks, materials and process conditions found on the actual line.

Cognex did not publish comparative customer results with the announcement, so manufacturers will need evidence from their own applications.

Country of origin becomes part of the design record

The UK government used the week to press for British inclusion in the European Union’s proposed Made in Europe policy.

The policy is intended to reduce reliance on Chinese components by giving preference to goods manufactured in Europe. The UK is seeking to prevent British suppliers from being excluded from European automotive, defence, technology and manufacturing programmes.

Around £15 billion of UK automotive exports are sold into European markets each year. The final rules have not been settled, so redesign decisions would be premature. The direction should still be reflected in sourcing records.

Country of origin is often collected after a part has been selected. Procurement preferences and requirements for local content can turn it into an early design constraint.

A qualified component may become commercially unsuitable for a particular contract. A change in fabrication, packaging or final assembly location could also alter how the finished product is treated.

The pattern beneath the announcements

AI infrastructure is no longer only increasing demand for advanced processors.

It is changing optical architecture, consuming memory capacity, raising power requirements and placing more pressure on production control. Governments are also taking a stronger interest in where components and assemblies originate.

The processor may establish the performance target. Memory availability, optical reach, power component life, manufacturing traceability and procurement eligibility increasingly decide whether the system can be built repeatedly and sold into its intended market.