Technology Transfer

In the spring of 2016, Lisa Su flew to China to seal a deal that would later embarrass her. AMD was a year and a half into her tenure as chief executive, the stock was climbing out of penny territory, and the company needed cash. China needed something else. For more than a decade Beijing’s planners had been chasing a domestic x86 server processor and getting nowhere. Intel and AMD held the only two licenses to the instruction set that mattered, locked behind a 1976 cross-licensing agreement that all but guaranteed neither would ever leak. AMD was prepared to bend the lock.

The structure her lawyers built was a small marvel of corporate origami. AMD did not sell its x86 designs to a Chinese company. Instead, it partnered with a Tianjin holding entity called Tianjin Haiguang Advanced Technology Investment, abbreviated to THATIC and pronounced as if it were a sneeze. Together they created two joint ventures. The first, Haiguang Microelectronics, would be 51 percent owned by AMD and hold the licensed intellectual property. The second, Chengdu Haiguang Integrated Circuit Design, soon known by the punchier name Hygon, would be majority-owned by THATIC and license the IP from the first venture to build chips for the Chinese server market. AMD took home $293 million plus royalties. The Chinese side took home the right to manufacture an x86 processor that the U.S. Department of Defense had spent forty years assuming China could not build. Su later told reporters that AMD had given the Chinese partners only the original Zen architecture, not the source-level register transfer logic, and that future generations would not be shared. The first sentence was true. The second became true, but only after Washington forced the issue.

This was tech transfer in the present tense. It was not espionage, not a midnight visit by a man with a thumb drive. It was a deal closed in conference rooms with bankers and IP lawyers, blessed in advance by an export-control regime that had not yet realized what it was looking at. By 2018 the joint venture had shipped the Hygon Dhyana, a chip whose silicon was, as the trade press tartly put it, indistinguishable from AMD’s Epyc except that the cryptographic instructions had been swapped from American standards to China’s SM2, SM3, and SM4 family. By June 2019, BIS had added Hygon, its sister entity Chengdu Haiguang Microelectronics, the server maker Sugon, and the Wuxi Jiangnan Institute of Computing Technology to the Entity List, citing their role in supercomputers with declared military end-uses. The first edition of the deal was the only edition that would ship.

The AMD-THATIC arrangement was the cleanest case of a pattern that ran through China’s semiconductor catch-up across the 2010s. Beijing had concluded after the failures of the 1990s and early 2000s that China could not, in any reasonable timeframe, originate leading-edge chip technology from inside its own laboratories. The country lacked the talent depth, the equipment base, and the iterative customer feedback that had compounded inside the Western and Taiwanese ecosystems for thirty years. So the State Council pursued a parallel strategy: industrial subsidies for domestic firms, equity acquisitions for foreign companies that would sell, and joint ventures and personnel recruitment for those that would not.

Each leg left a different footprint. The acquisitions were public, dramatic, and politically combustible. Tsinghua Unigroup’s $23 billion bid for Micron in mid-2015 collapsed within weeks under CFIUS attention, the highest-profile of a string of failed deals that belongs to a later chapter. The joint ventures and the personnel flows belong here. They were quieter and harder to police, and they did most of the work.

Inspur, China’s largest server maker, took a different route to the same destination. By the early 2010s it was selling its TS K1 mainframe-class machine into Chinese banks and state-owned enterprises, fighting IBM head-on. The two companies traded jabs through 2013. Inspur ran a marketing campaign called I2I, for IBM-to-Inspur, which leaked as a deliberate signal that the country’s banks were swapping out American iron for Chinese boxes. Then, in August 2014, IBM announced a partnership. Both companies would make IBM’s DB2 database and WebSphere application server software run on the Inspur K1. IBM would supply POWER8 processors and design assistance. Inspur would join the OpenPOWER Foundation, the open consortium IBM had launched a year earlier to share the POWER instruction set with anyone willing to build hardware around it.

OpenPOWER was, on paper, a counter-Intel coalition. IBM had watched Power’s market share decay through the 2000s and concluded that the only way to keep its architecture alive was to give it away. By 2014 Inspur had joined Google, Nvidia, Mellanox, and Tyan; within a year a dozen Chinese firms had clustered into a China POWER Technology Alliance to coordinate Chinese implementations. From IBM’s vantage, the trade was rational: the Chinese market was closing to foreign servers, and the only way to keep any Power chips running in China was to license them to a local partner. From Beijing’s vantage, it was a bargain. China’s planners had never managed to build a competitive server CPU; OpenPOWER handed them the design language for one.

Qualcomm walked into a similar deal a year later, with similar motivations and a worse outcome. By 2016 Qualcomm was sitting under a Chinese antitrust judgment that had cost it almost a billion dollars and forced the company to renegotiate smartphone royalty terms with every Chinese handset maker. It also wanted into server chips, a market dominated by Intel and resistant to ARM-based challengers. The Guizhou provincial government wanted a flagship semiconductor venture. In January 2016 the two announced Huaxintong Semiconductor Technology, HXT, with combined investment that would reach $570 million. Qualcomm would contribute its ARM-based Centriq server core. Guizhou would supply the funds, the data-center demand, and the political cover. The first chip, the StarDragon 4800, taped out and reached limited production in late 2018. By April 2019 the venture was being wound down. Qualcomm had exited the ARM server business globally; Guizhou had no appetite to keep funding a Chinese clone of an obsolete strategy. The CEO resigned. The designs went into a drawer.

Each of these joint ventures had the same shape. An American company facing pressure in China, antitrust or market-access or competitive, agreed to share design IP in exchange for partial ownership of a Chinese-domiciled entity selling into the protected mainland market. Each was structured to comply with the letter of the relevant export-control regime in 2014 or 2016. Each became, within three to five years, a target of the same regime once the political climate shifted. The ventures were not violations so much as a stress test of how much the rules had assumed away. CFIUS, as the AMD case showed, had not historically claimed jurisdiction over joint ventures, only over inbound acquisitions. The Entity List, before 2019, was applied sparingly. The whole architecture of U.S. export controls had been built for a Cold War in which the adversary’s economy was sealed off from the world’s, not for one in which American firms could be paid to walk capability across the border in plain daylight.

The personnel flows were quieter and, in the long run, more consequential. Chips are designed by people, and the people who knew how to design them in 2015 were concentrated in a few thousand engineering offices in Hsinchu, San Jose, Hillsboro, and Suwon. The most accessible of those reservoirs, from Beijing’s point of view, was Taiwan. Same language, easy travel, deep family networks across the strait, forty years of process expertise inside TSMC, UMC, MediaTek, and a long tail of smaller firms.

Charles Kau, by 2015, had become the public face of how the flow worked. Born in Taiwan and trained in chemical engineering at North Carolina State, Kau had worked at Fairchild Semiconductor in California in the 1970s, returned to Taiwan, helped found the flash-memory firm Macronix, and later led Nanya Technology and Inotera. The Taiwanese press called him the godfather of the island’s DRAM industry. In late 2015, at sixty-one, he resigned from Inotera and took a role as global executive vice president of Tsinghua Unigroup, the same Beijing-backed conglomerate that had just had its bid for Micron blocked. Industry observers called it a defection. Kau called it an offer he could not refuse. His tenure in Beijing did not last long; by 2018 he had stepped down. The signal had been sent. The most senior memory executive in Taiwan had crossed the strait, and he was not the only one.

Below the level of celebrity hires, the recruitment was systematic. Programs like the Thousand Talents Plan, established in 2008, offered overseas Chinese researchers cash bonuses, subsidized housing, and lab funding well in excess of what Western academic institutions could match. The semiconductor-specific recruitment targeted ethnic Chinese engineers in Silicon Valley and Hsinchu. By the late 2010s, Taiwanese authorities estimated that Chinese firms had drawn nearly sixty thousand professionals from across the diaspora since 2008, of whom several thousand worked on chip design. Recruiters offered three to five times the Taiwanese rate, plus relocation, plus signing bonuses tied to patent counts or yield improvements.

The most aggressive recruiters did not even bother with relocation. They built shadow companies inside Taiwan itself. In March 2021, Taiwanese prosecutors raided two firms in New Taipei, WiseCore Technology and IC Link, and detained nineteen people. The companies, on paper, were independent Taiwanese chip-design houses. In practice, they had been set up by Bitmain, the Beijing-based maker of Bitcoin-mining chips, to recruit engineers away from TSMC, MediaTek, and the local AI-chip firms without ever asking those engineers to leave the island. The engineers worked from Taiwan, designed chips ostensibly for Taiwanese customers, and shipped the resulting designs to Bitmain’s affiliates in mainland China. Researchers at Taiwan’s Research Institute for Democracy, Society, and Emerging Technology named the practice the remote-poaching model. Bitmain’s affiliates had absorbed years of work on edge-AI inference silicon by the time the offices were raided.

The TSMC outflow had a separate and earlier history. In late 2016, Chiang Shang-yi, who had served as TSMC’s chief technology officer through the foundry’s leap from 130-nanometer copper to 65-nanometer and beyond, joined the board of SMIC as an independent director. Three years later, in mid-2019, he agreed to become chief executive of Wuhan Hongxin Semiconductor Manufacturing, a $20 billion startup whose principal achievement turned out to be having been built. The fab broke ground, hired engineers, leased one of the few remaining ASML lithography systems available outside the U.S.-allied bloc, and then collapsed amid local-government recriminations and unpaid contractors. Chiang resigned in mid-2020 and rejoined SMIC as vice chairman that December. Three days after he started, the United States added SMIC to the Entity List. Chiang, a U.S. citizen as well as a Taiwanese national, found himself unable to engage with the very process technology he had been hired to advise. He left in 2021 and gave an interview in which he called his decision to join SMIC the most foolish of his life.

Liang Mong-song, who had been Chiang’s protégé at TSMC and who had similarly left the company after being passed over for promotion, took a more durable route. After a stint at Samsung that ended in a Taiwanese trade-secrets lawsuit and a court order forbidding him from working for the Korean firm past the end of 2015, Liang joined SMIC as co-chief executive in late 2017. He brought with him, by industry estimates, several dozen former TSMC engineers. By 2019, SMIC was sampling 14-nanometer process chips. By 2022, internal sources confirmed it had reached a 7-nanometer node, several years ahead of the trajectory most Western analysts had projected. None of the silicon walked across the strait. None of the manuals did either. The people walked across the strait, and the silicon followed.

Beijing did not invent any of this. The mechanisms of joint venture, licensing, and recruitment had been the way every Asian latecomer had broken into chips since the 1960s. Texas Instruments had assembled its first transistors in Taiwan in 1968, training a generation of local engineers. Samsung had purchased a 64-kilobit DRAM design from Micron in 1985, hired American engineers willing to relocate to Suwon, and entered the memory market. TSMC had been seeded by Morris Chang’s connections to Texas Instruments and General Instrument, and had taken its first leading-edge process node from a license with Philips. The pattern of foreign-trained engineers returning home, of joint ventures gradually building local capability, of technology percolating outward through the diaspora, was older than the chip itself.

What was new in the China case was the scale, the speed, and the political reading. Patrick McGee’s account of Apple’s two-decade investment in Chinese manufacturing, published in early 2025, captured a parallel mechanism in consumer electronics. Apple, McGee documented, had spent on the order of $275 billion upgrading its Chinese suppliers between 2016 and 2021, peaking at perhaps $55 billion a year, and had trained more than 28 million Chinese workers since 2008, more than the entire labor force of California. McGee’s framing was that this had not been theft or coercion or even forced technology transfer in the customary sense. It had been a transfer of process knowledge in exchange for the lowest unit costs and fastest ramp times the world had ever seen, conducted enthusiastically by an American company that did not realize, until much later, what it was teaching the host country to do without it.

The semiconductor industry’s version had a heavier overlay of national security from the start, but the underlying physics were similar. The knowledge that mattered, in chips as in iPhones, was not the patent. It was the operational tacit knowledge held by engineers who knew how to coax a yield curve out of a finicky etcher, who knew the failure modes of a specific photoresist on a specific cluster tool. That kind of knowledge moved on airplanes, in boardrooms, and in offer letters, not in PDF files. The 2018 USTR Section 301 report, which formally labeled China’s joint-venture requirements as a system of forced technology transfer and put the cumulative cost to the U.S. economy at fifty billion dollars a year, made the case in legal language. The day-to-day reality was that the cost was being paid voluntarily, transaction by transaction, by American companies that had decided the China market was worth more than the marginal leakage of any single deal.

Doug Fuller, in his 2016 study of China’s tech ecosystem, argued that the country’s most successful semiconductor companies were neither the state-owned national champions nor the pure multinationals but a category he called hybrid firms, run by ethnic Chinese managers with foreign-company experience and financed by a mix of foreign venture capital and Chinese state money. Hygon, Inspur, and HXT fit Fuller’s description in their corporate skeletons. Bitmain’s remote-poaching network fit it in personnel. SMIC fit it most exactly of all: founded by a U.S.-trained Taiwanese executive, capitalized first by Western VCs and later by the Chinese state, staffed by engineers drawn from TSMC and Samsung, and operating in legal gray zones that depended on which border one was looking across.

Through the back half of the 2010s, that hybrid model produced enough domestic capability to alarm the Pentagon and enough leakage to enrage Congress. By 2019 the Entity List was being used to claw back, retroactively, what the joint ventures of 2014 and 2016 had handed forward. AMD was barred from licensing further generations of Zen to its Chinese partners. IBM was discouraged from extending OpenPOWER cooperation. Qualcomm exited Guizhou. The flows of engineers across the Taiwan Strait did not stop. The remote-poaching networks adapted. The pattern shifted from joint ventures to talent acquisition, and from talent acquisition to whatever quieter mechanism would replace it.

Inside Beijing, the planners drew the lesson that the channels could be closed one by one. They began preparing for a different kind of catch-up, one done without American partners willing to sit in conference rooms in Tianjin. The window of joint ventures and silent personnel flows had opened in 2014 and would, in spirit, close around 2020. By the time it shut, an enormous amount had moved through it, and almost none of it had moved illegally.