NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

The Chinese Machine That Stacks AI Memory Says It Passed a Customer Test

Shenyang-based Piotech says its wafer-to-wafer hybrid bonding tool cleared customer validation and is buying a PVD maker to become a one-stop shop for 3D chip integration — the unglamorous layer under the AI memory boom.

2026-11-19 · 1182 words · NeuroAI
The Chinese Machine That Stacks AI Memory Says It Passed a Customer Test

A chip stack is not held together by glue, solder or hope. It is held together by two polished surfaces, pressed into contact until copper meets copper and the atoms stop noticing the seam.

The machine that does this is called a hybrid bonder. It is not a lithography scanner, it does not appear in export-control headlines, and it is becoming central to the AI supply chain, because high-bandwidth memory and advanced packaging are both stacks.

On 28 October, a Chinese supplier of that class of tool said its newest bonder had cleared customer validation.

Key takeaways

  • Piotech (拓荆科技), a Shenyang-based semiconductor equipment maker, said its Dione 300 eX wafer-to-wafer hybrid bonding (混合键合) tool passed customer validation, claiming performance at the level of the most advanced international products.
  • The company shipped more than 4,400 process chambers cumulatively and reported RMB 5.22 billion (≈ US$735M) in revenue for the first nine months of 2026, up 23.7%, with net profit up 336.8%.
  • It has agreed to buy 82.97% of Wuxi Shangji, a PVD and dry-etch toolmaker, for RMB 2.5 billion (≈ US$352M), adding the one major deposition category it lacked.
  • Hybrid bonding is the process behind stacked image sensors, 3D NAND, and increasingly HBM and chiplet packaging — which is why a mid-sized equipment firm matters to the AI build-out.

Why bonding is the quiet bottleneck

For two decades, chipmakers connected dies with solder bumps: place a die, reflow, and the joint is 20 to 40 micrometres wide. That works until you need thousands of connections per square millimetre and the bumps bring too much resistance, heat and space.

Hybrid bonding removes the solder. Two wafers are planarised until they are flat to within nanometres, aligned, and pressed. The dielectric surfaces bond at room temperature; a later anneal makes the copper pads fuse. Interconnects shrink to a few micrometres, and the stack behaves almost like one piece of silicon.

The hard part is not the concept. It is alignment: you position two wafers with hundreds of thousands of pads, and a fraction of a micron of drift ruins the die. Then there are particles — one speck between the surfaces creates a void that propagates across the pair — and throughput, because a bonder taking ten minutes per wafer cannot serve a memory line.

This is why so few companies make these tools. Validation by a real customer matters more than any specification sheet.

What Piotech actually said

In its third-quarter report, released on the evening of October 28, Piotech said the Dione 300 eX had passed customer validation, using more advanced alignment technology with higher alignment accuracy, bonding accuracy and throughput. The company's wording was that the product can reach the level of the most advanced comparable international products — a company claim, not an independently measured result, and no customer was named.

The rest of the report is a reminder that this is not a startup. Revenue for the first three quarters was RMB 5.22 billion (≈ US$735M), up 23.7% year on year. Net profit attributable to shareholders was RMB 2.43 billion (≈ US$342M), up 336.8%, though that figure is flattered by non-operating items; the non-recurring measure was RMB 929 million (≈ US$131M). Cumulative chamber shipments passed 4,400 units, across memory, logic, packaging, power devices, Micro-OLED, silicon photonics and image sensors.

The acquisition is the real strategy

Earlier, on 15 October, Piotech had announced a plan to acquire 82.97% of Wuxi Shangji for RMB 2.5 billion (≈ US$352M) in shares and cash, with up to RMB 1.82 billion (≈ US$256M) in accompanying fundraising.

Wuxi Shangji makes physical vapour deposition equipment — the sputtering tools that lay down metal and barrier films — plus dry etch. That fills the one hole in Piotech's portfolio: the company is strong in PECVD, ALD and other thin-film deposition (薄膜沉积) steps, but has had no PVD product of its own.

Put the pieces together and the intent is clear. Piotech wants to sell a matched set for 3D integration: deposit the films, etch the vias, bond the wafers, inspect the result. Memory makers would rather qualify one partner than four, because every vendor is another set of recipes and another yield argument.

The targets attached to the deal are modest: Wuxi Shangji's sellers committed to revenue of at least RMB 350 million, 450 million and 610 million for 2026 through 2028. This is a capability purchase, not a revenue purchase.

The demand signal underneath

The context is a memory market being pulled upward by AI. Industry forecasts cited by Chinese memory module maker Biwin put Q3 2026 DRAM contract prices up 13% to 18% and NAND up 10% to 15%.

Rising prices push memory makers to add capacity and upgrade process. Each shrink makes capacitor holes deeper and narrower and metal fill harder, which adds etch and deposition steps per wafer — and advanced memory is where Piotech says its film tools are being adopted fastest.

Then there is packaging. HBM is a stack of DRAM dies connected by through-silicon vias and bonded to a logic base die. Every generation increases the number of dies and tightens the pitch, and CoWoS-class packaging, chiplets and 3D NAND string stacking all push the same way: more bonding, more often, at finer pitch.

That is the market Piotech is positioning for: not leading-edge logic, but the stacking layer AI accelerators cannot do without.

Honest limitations

This article is based on Piotech's third-quarter report and its acquisition announcement as reported by Chinese securities media. I have not seen the customer validation data, the identity of the qualifying customer, the tool's measured overlay accuracy, throughput or defect density, or any independent test of the "most advanced international level" claim.

That claim is the company's own. Public specifications for hybrid bonders are sparse, so even a careful reader cannot verify the ranking.

The Wuxi Shangji acquisition had been announced but not completed at the time of writing. Revenue commitments are targets agreed by sellers, not forecasts by the buyer, and the deal still requires regulatory and shareholder approvals.

Finally, the 336.8% net profit growth includes non-operating gains; the non-recurring figure is the better guide to the operating business.

What readers can do now

  • If you track the AI supply chain, add bonding to your map. Lithography and HBM capacity get the headlines, but hybrid bonding throughput is a real constraint on how fast stacked memory and chiplets can scale.
  • If you are evaluating Chinese equipment claims, separate "validated" from "adopted". Passing a customer's qualification is a gate; volume orders, repeat orders and yield data are the evidence that follows.
  • If you invest, read the non-recurring line. Equipment companies in an upcycle often report net profit inflated by fair-value gains and share disposals. The operating number tells you what the tools actually earn.

The bonder will never get a launch event. It sits in a cleanroom, aligns two wafers, and presses. But the next generation of memory and packaging runs through machines like it — and the country that can build the machine has a say in how fast the stack gets taller.

Related coverage

Cite this article

NeuroAI (脑机网), "The Chinese Machine That Stacks AI Memory Says It Passed a Customer Test," 19 November 2026, https://neuroai.site/a/na-chips-piotech-hybrid-bonding

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Chips & Compute” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories