The cheapest operation in computing has always been the one you do not have to heat up. A small Beijing alliance is betting that the next leap in AI efficiency comes from replacing part of the electrical pathway inside a server with light.
Three companies with very different specialties — a chip startup, a GPU maker, and a large-model lab — have decided to build a shared compute base rather than compete for the same slice of the stack. The wager is that photons, not just electrons, can carry the load of inference.
The three players and why they fit
On July 8, 2026, in Beijing, MakeSens (每刻深思) signed a strategic cooperation agreement with Biren Technology (壁仞科技) and StepFun (阶跃星辰), and the three jointly released a photoelectric-fusion AI compute platform. The event carried the tagline "optical computing power, heterogeneous coexistence (光算聚力·异构共生)."
The fit is complementary rather than overlapping:
- MakeSens contributes spatial optical computing chips (空间光计算芯片), born from Tsinghua University research and founded in 2020, with a processor it calls the OPU that it says breaks the bottleneck of Moore's Law.
- Biren Technology supplies domestic general-purpose computing power (算力) through its flagship GPU line, plus cluster scheduling and super-node systems needed to run large models at scale.
- StepFun handles large-model optimization, operator-level adaptation, and scenario tuning, closing the gap between algorithms and heterogeneous hardware.
Each partner brings the piece the others lack: light-speed inference, a domestic GPU floor, and a model to run on top.
What "photoelectric fusion" actually means
The phrase sounds abstract, but the mechanism is concrete. Today's servers move and process data with electricity, and electronic chips are bumping against physical limits — their energy efficiency gains are flattening as transistors approach their smallest useful size.
Photons travel faster, use less power, and can compute across many paths in parallel. The alliance's idea is to keep electricity where it is strongest and hand the inference workload to optical components where light wins. The stated full stack is "optical computing inference chip plus general GPU plus large-model algorithm," a deliberately heterogeneous design where each layer does what it does best.
Qiao Fei (乔飞), a Tsinghua associate professor and MakeSens co-founder, put the industry's honest problem plainly: optical technology is verifiable in the lab but historically hard to scale into mass production. The partnership is framed as an attempt to close exactly that gap.
The product and its claimed first
Alongside the agreement, the three unveiled what they describe as the first optical-computing-plus-general-electric-computing AI compute product. That "first" is the partners' own description, and it should be read as a claim rather than an independently verified record.
What is less contestable is the architecture's intent. By binding a spatial optical chip, a domestic GPU base, and a large model into one tuned system, the group is trying to turn a research direction into a shippable server-class product. The next steps the companies named are joint system debugging, industry-scenario pilots, and continued work on the software-hardware toolchain.
The alliance was witnessed by the Beijing-Tianjin-Hebei National Technology Innovation Center, whose director Wang Dixiang (王茤祥) attended, with ecosystem partners including CITIC Securities (中信证券) also present — a signal that the effort has institutional and financial backing beyond the three firms themselves.
Why this matters for the compute shortage
The strategic pitch is twofold. First, lower the deployment cost of large models; second, raise response efficiency across both edge and cloud scenarios. If optical inference can shave the energy bill of serving a model, it attacks the single biggest constraint on AI adoption: the cost of running intelligence after it is trained.
For China specifically, the alliance is also a story about domestic computing power (算力) autonomy. By combining a homegrown GPU path with a distinct optical approach, the partners are diversifying the supply of compute away from any single technology or vendor. Biren's representatives framed domestic heterogeneous compute as the core foundation of the AI industry, and StepFun's side emphasized native coordination between chips and models.
Trend reading
According to Li An, chief scientist at BrainNet (脑机网), photonic inference is promising for specific workloads but the real test will be whether the toolchain matures enough for ordinary enterprises to adopt it without a physics doctorate on staff.
Honest limitations
The "world's first" product claim is the partners' own description and has not been independently benchmarked; optical computing remains closer to pilot than to mass deployment. No public performance, price, or shipment figures were released, so cost and energy savings are assertions rather than measured results. The alliance depends on three independent roadmaps staying aligned, and on domestic supply chains for both optical and GPU components, either of which could stall the timeline.
What readers can do now
- Cloud and inference buyers should track the trio's pilot results as a possible low-power alternative for specific model-serving workloads, especially where energy cost dominates the bill.
- Domestic GPU and chip firms should watch whether optical inference becomes a complementary layer rather than a replacement, and plan their integration roadmaps accordingly.
- Investors should treat the photoelectric-fusion narrative as early-stage and demand published benchmarks before reading it as a near-term fix for the compute shortage.