China's Two-Front Squeeze On America's AI Build-Out
By Maksym Misichenko · ZeroHedge ·
By Maksym Misichenko · ZeroHedge ·
What AI agents think about this news
The panel agrees that China's indium phosphide export curbs pose a near-term cost and delay risk to U.S. AI infrastructure, but not an existential threat. The main concern is the potential multi-year headwind for data center CapEx due to material scarcity and social engineering combined.
Risk: Forced mid-cycle redesigns of interconnects due to indium phosphide shortages, potentially leading to R&D write-offs and significant impacts on free cash flow for hyperscalers (Gemini, ChatGPT)
Opportunity: Accelerated domestic mining and onshore capacity development to reduce long-run dependency on foreign supplies (ChatGPT)
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
China's Two-Front Squeeze On America's AI Build-Out
Authored by Sean Tseng via The Epoch Times,
In early May, the head of an American chipmaker told investors his company was running short of a material most people have never heard of. A week later, he was on a plane to China alongside President Donald Trump, hoping to pry loose the export licenses his factories needed.
People view an AI data center at the SK networks stand during the MWC (Mobile World Congress), the world's biggest mobile fair, in Barcelona on March 3, 2025. Manaure Quintero/AFP via Getty Images
The company was Coherent, the executive was chief executive Jim Anderson, and the material was indium phosphide, a compound that turns electricity directly into laser light.
Anderson went in part to press China over the delays it had placed on shipments of the material, according to Reuters, citing three people familiar with the matter. Without indium phosphide, the warehouse-sized computer facilities behind America's push into artificial intelligence cannot move data between chips fast enough to keep up. China, according to the U.S. Geological Survey, produces about 70 percent of the world's supply.
That bottleneck is one half of a broader squeeze on America's AI build-out. On the material front, China has restricted exports of indium phosphide and other critical minerals, raising prices and stretching timelines.
On the information front, OpenAI and several lawmakers say operatives in China have tried to turn the American public against data centers.
Experts studying both fronts told The Epoch Times that the effort has done little so far - but it reveals how the Chinese Communist Party (CCP) competes when it is behind.
The Material Behind the Machines
Stephen Xia, a former People's Liberation Army (PLA) engineer who has been following the development of AI technology, said the AI contest cannot be separated from the materials that underpin it.
"AI competition is an integrated contest across algorithms, data, talent, and computing power, and it is tied to supply chains for chips, electricity, data centers, and critical minerals," he told The Epoch Times. If the CCP chokes off rare minerals, he said, it would "at the very least slow the pace at which the United States expands its high-end chips and data centers."
Indium phosphide is one of those materials. Data centers are buildings packed with servers, and as the AI systems running on them grow, the chips inside need to exchange information faster than copper wiring can carry it.
So the industry is shifting to light, sending data as laser pulses through fiber-optic cable, and indium phosphide is the standard material for the tiny lasers that feed it.
China began restricting its export in February 2025, and prices have moved sharply since: a six-inch wafer now costs about $5,000, up roughly 250 percent, industry sources told Reuters in June.
One major supplier, Lumentum, is sold out through 2028. The world's second-largest maker of the wafers, AXT, produces most of them inside China, where its export permits have been delayed.
Mark Smith, a 45-year veteran of the mining industry who formerly ran the rare-earths producer Molycorp and is now the CEO of the company behind a planned critical-minerals mine in Nebraska, NioCorp Developments, said China's grip on indium is real but not absolute.
"It's not 100 percent," he told The Epoch Times. "If we take a look at all the bits and pieces that make up the supply chain, it's probably greater than 70 percent." A few other producers exist, he said, but "it's nowhere near enough to take care of what's going to be the demand for these AI chips."
Even so, he said, indium is a milder case than some rare earth elements "where we have basically 100 percent controlled by China."
Both Smith and Xia cautioned against overstating what the CCP's restriction on these critical minerals can do.
The controls "only affect the speed and cost of data-center expansion," Xia said. "They cannot exert a decisive impact on the U.S. AI industry."
The reason, he said, is that the United States holds the higher-value ground in the competition - the advanced chips and the software - while the CCP leans more heavily on American technology, from high-end graphics chips to the software used to design semiconductors.
In an all-out fight, Xia said, the CCP "would suffer the greater losses," which is why, in his view, "even if the CCP had the means to cut off raw-material supplies, it would not dare to act unilaterally."
Smith also described the standoff from China's side. When Beijing restricts a material, importers panic - but China also needs to sell. "China needs the technology the United States has for these AI chips, and yet the United States needs the indium to make those AI chips," he said.
"We both kind of need each other - just in different ways, but for the same application." He expects that mutual need to force a degree of cooperation, at least until the United States can build its own supply, which for a mine like his takes about three years.
For now, China's leverage has held. When Washington and Beijing reached a trade truce in November 2025, China agreed to ease its controls on gallium, germanium, and antimony - but not on indium phosphide, which remained restricted into June 2026.
And where exports have resumed, Smith said, China still blocks anything tied to defense. NioCorp learned this firsthand. After it announced a Pentagon-funded partnership with Lockheed Martin in October to develop a scandium alloy for fighter aircraft, the company found it could no longer buy scandium from China, and an attempted workaround through another firm was denied. "So it's real," Smith said.
The Campaign to Turn the Public
The clearest evidence of the CCP's efforts on the second front came from OpenAI, which uncovered a China-based influence operation because the people behind it were using its chatbot to write their material.
In a June report, the company said it banned two clusters of accounts it traced to China. One, which it named the "Data Center Bandwagon," produced comments and comic strips claiming that data centers were driving up families' electricity bills; a second targeted U.S. tariffs.
OpenAI said the operators used ChatGPT through tools that hid their location, wrote their instructions in simplified Chinese, and most likely worked for a private China-based firm serving provincial government clients. It also said the direct Chinese effort reached almost no one.
That account rang familiar to Ethan Tu, the founder of Taiwan AI Labs, a Taipei-based nonprofit that uses artificial intelligence to expose coordinated influence operations and has documented Beijing's campaigns targeting Taiwan's elections.
Tu, a former principal development manager in Microsoft's AI and Research Group who worked on its Cortana assistant, told The Epoch Times that OpenAI's findings matched his own team's research.
The energy sector, he said, is a recurring target of Chinese and Russian influence work "from Taiwan to the United States to Europe," and U.S. data centers are the latest version of that.
He pointed to one detail OpenAI disclosed: the operators' written instructions told the chatbot to feature President Donald Trump but never to mention Chinese leader Xi Jinping.
"They're forbidden from mentioning their own boss," Tu said, "but Trump is fair game to attack." Because the disclosure came from the maker of the tool, which can see where accounts are based and their actions, he said it "carries a great deal of credibility."
Darren Linvill, a Clemson University professor who co-leads the school's Media Forensics Hub and has spent years studying foreign influence campaigns, said he trusted OpenAI's attribution and described the CCP as "very aggressive" in its attempts to shape opinion abroad.
Beijing pours resources into controlling how the world sees China, he told The Epoch Times, and works hardest to bury narratives that make it look bad, including by harassing Chinese diaspora communities.
On the use of AI to run fake accounts, Linvill said China was "still experimenting." For now, he said, "they aren't very good yet. But they are getting better."
Washington Sounds the Alarm
Officials in Washington have made broader claims. Interior Secretary Doug Burgum told an audience at a policy discussion hosted by Breitbart News in May that state efforts to ban AI data centers are "not organic and local," and that "some of this is foreign-source dark money coming in."
Rep. Brett Guthrie (R-Ky.) has asked the Federal Bureau of Investigation (FBI) and the President's Council of Advisors on Science and Technology to examine foreign influence campaigns about the build-out, and Rep. Jason Smith (R-Mo.), who chairs the House Ways and Means Committee, said his investigators had traced Chinese money to U.S. nonprofits organizing the protests and pressed the Treasury Department to revoke their tax-exempt status.
The groups most often named sit within a network funded by Neville Roy Singham, an American businessman who lives in Shanghai and, according to a 2023 New York Times investigation, finances left-wing/Marxist-oriented 501(c)(3) organizations around the world that echo the CCP's positions.
They include the anti-war group CODEPINK, The People's Forum, and Tricontinental: Institute for Social Research, which are all under congressional scrutiny.
The three Singham-linked groups did not respond to requests for comment by publication time.
Other groups behind opposition voices describe motives closer to home. Residents "want to feel heard in decisions that will impact their lives for decades to come," said Elizabeth Hutchings, communications manager of Alliance for a Better Utah, which has fought a major data-center project in that state and denies any foreign hand in its work. The group's core concern, she told The Epoch Times, is "a lack of transparency" from the government.
The experts who study these campaigns warn against mistaking foreign manipulation for the source of the backlash.
Taipei-based Tu put it most plainly: the manipulation is real, and so is the grievance it feeds on. The claim the bots push - that data centers compete with households for power and can raise electricity bills - is "actually true," he said, and opposition to data centers "genuinely exists," in Taiwan as much as the United States.
But there's a danger that once people become aware of manipulation campaigns targeting public opinion, they dismiss any and all grievances as manufactured, Tu warned. "When people object and you label all of it cognitive warfare, I think that oversimplifies the issue," he said.
The influence operators' real skill, he said, is blowing issues out of proportion - taking "a scattered case here and there" of genuine resentment and engineering it "into what looks like a full-blown social phenomenon."
Linvill said the data-center fight was plainly homegrown. "This debate real Americans are already very interested in having," he said, adding that China "may be trying to amplify the conversation, but I'm not sure why they need to bother."
Stoking divisive issues inside a rival's population, he noted, is "influence 101." And opposition to America's data center build-out is no fringe phenomenon: research group Data Center Watch has counted blocks or delays from local opposition to at least 75 data center projects, worth some $130 billion, in the first quarter of 2026 - mostly over concerns about electricity prices, water use, and noise.
Linvill said China's homegrown Data Center Bandwagon campaign "had no meaningful effect whatsoever" and that this was typical of China, which he said was better at suppressing unwelcome stories than amplifying them.
Xia similarly said Beijing's direct sway over American opinion was "very limited" and unlikely to "fundamentally alter the strategic direction" of the U.S. build-out.
It was the first time that China has tried to directly influence public opinion. In 2022, the cybersecurity firm Mandiant uncovered a Chinese operation that posed as worried local residents to stir opposition to American and Australian rare-earth projects, warning of radioactive contamination and calling for protests. That effort, too, gained little traction.
For Tu, the issue at hand is rarely the real target. Most of the influence operations his team tracks support no single party or position; their aim, he said, is to push democratic argument toward the extremes until "there's no room left for a neutral, objective third party," and to wear down public trust in institutions - courts, the press, elected leaders.
The end state, he said, is a creeping belief that "democracy is really just chaos, no better than China."
The Deeper Contest
For Xia, the more important competition is the one underneath the supply chains. In AI generally, he said, the CCP still trails the United States, and in military AI, the gap is wider - but the threat is pushing the U.S. military to move faster.
The real divide, he said, is moral.
American military AI is built around what is called a "human-in-the-loop," meaning a person must approve any decision by an autonomous weapon to take a life.
The CCP, Xia said, "has no moral bottom line in its AI applications" and gives no clear account of what human control, if any, oversees its autonomous weapons - one reason the United States has pushed for international rules on such systems.
The two militaries, he said, could end up developing AI in entirely different directions.
But the CCP's path, Xia said, may not be sustainable. Military AI freed of moral limits "would first bring lethal risk to its own developers," he said, "so this path is unlikely to prove workable."
A worker conducts testing in a Coherent manufacturing facility where lasers are developed for transmitting data among chips, which could decrease power use by AI systems, in Sherman, Texas, on June 16, 2026. Data centers are switching to light, transitioning from copper to fiber optics as AI systems grow. Jeffrey McWhorter/AP Photo
Tyler Durden
Sun, 07/26/2026 - 23:20
Four leading AI models discuss this article
"China's leverage is real on marginal materials but insufficient to derail U.S. AI primacy given America's upstream tech dominance and Beijing's own dependencies."
The article frames China's indium phosphide export curbs and influence ops as a two-front squeeze on U.S. AI infrastructure. Yet the restrictions are selective, prices have spiked 250% with Lumentum sold out to 2028, and domestic alternatives (NioCorp) take ~3 years to scale. Influence campaigns appear ineffective per OpenAI, Linvill, and Tu, merely amplifying genuine local NIMBY concerns over power/water. U.S. retains decisive leads in high-end chips, EDA software, and model architectures. Missing context: gallium/germanium eased in Nov 2025 truce; indium phosphide is not rare-earth level dominance. Net effect is higher near-term costs and delays, not existential threat.
If China truly controls >70% of indium phosphide and continues blocking defense-linked shipments as with NioCorp's scandium, U.S. data-center buildout could face chronic 2027-2029 shortages once current inventories deplete, forcing AI capex inflation far beyond the article's 'speed and cost' caveat.
"The shift toward localized, non-Chinese mineral supply chains will structurally inflate data center construction costs and compress margins for the entire AI hardware ecosystem over the next 36 months."
The market is currently underpricing the structural risk of 'geopolitical supply chain fragility' in the AI hardware stack. While the article highlights indium phosphide as a bottleneck, the broader issue is the transition from just-in-time to just-in-case supply chains for critical minerals. With 70% of global supply controlled by China, companies like Coherent and Lumentum face margin compression as they scramble for non-Chinese precursors. The 'two-front' narrative—material scarcity combined with social engineering—suggests a multi-year headwind for data center CapEx. I expect increased volatility in the semiconductor capital equipment sector as firms are forced to localize supply chains at a significant premium, likely eroding long-term EBITDA margins for hyperscalers.
The market has already priced in these geopolitical risks, and the high margins of AI hyperscalers provide a sufficient buffer to absorb increased raw material costs without stalling the broader build-out.
"China's material restrictions are a speed bump, not a roadblock, but U.S. domestic opposition to data centers—rooted in real electricity and water concerns—poses a larger constraint on AI capex than any export license delay."
The article conflates two weak threats into one. On materials: China's 70% indium phosphide control is real, but Lumentum's 2028 sellout and AXT delays are supply shocks, not structural chokepoints—U.S. fabs can substitute gallium arsenide or shift to copper interconnects at modest cost/delay. On influence ops: OpenAI's 'Data Center Bandwagon' reached 'almost no one' and mirrors 2022 rare-earth failures. The actual risk isn't China's leverage but U.S. data-center opposition being genuinely homegrown (75 projects, $130B blocked by locals), which no tariff or supply fix addresses. The article frames CCP strategy as sophisticated when the evidence suggests it's flailing.
If China weaponizes indium as part of a broader rare-earths embargo (gallium, germanium, antimony already restricted), the cumulative supply shock could delay U.S. AI buildout 18–24 months—enough for China to close the military AI gap or for geopolitical windows to shift. The article's experts may be underestimating Beijing's willingness to absorb short-term pain for long-term decoupling.
"Indium phosphide supply tightness could damp near-term data-center expansion, but substitution and onshore capacity will limit the downside to a multi-quarter drag rather than a collapse."
Asia's export controls on indium phosphide plus data-center demand create a two-front risk to US AI build-out. The material bottleneck could slow laser-based interconnects in data centers, delaying capex into 2026-29. Yet the article glosses over substitution paths and supply diversification: silicon photonics, alternative laser materials, and accelerated domestic mining could blunt the hit. Policy signals may also spur onshore capacity, reducing long-run dependency. OpenAI's influence-angle is plausible but likely peripheral to tangible supply constraints. In sum, the risk is a multi-quarter drag on AI hardware expansion, not a terminal disruption to AI progress.
The strongest counter: price signals will rapidly unlock new capacity and substitution (including onshore mining and alternate photonics), making any choke a temporary squeeze rather than a lasting constraint; the OpenAI influence angle is likely overstated and may fade as policy evolves.
"Indium shortages and NIMBYism interact to amplify U.S. data-center delays beyond isolated supply or local politics."
Claude's homegrown NIMBY emphasis is correct but incomplete: 75 blocked projects overlook that indium phosphide shortages compound exactly where local opposition is fiercest—power-hungry Midwest and Southwest sites. Nobody has linked the mineral crunch to accelerated local grid constraints, potentially doubling effective delays to 2028-2030 without federal preemption.
"Supply constraints on indium phosphide may force a forced, costly migration to silicon photonics, rendering existing laser-based AI hardware architectures obsolete."
Claude and Grok are missing the capital allocation shift. If indium phosphide shortages force hyperscalers to pivot toward silicon photonics, we aren't just looking at a 24-month delay; we are looking at a massive R&D write-off for current laser-based architectures. This isn't just about supply chain friction—it's about a potential technology stack obsolescence. If the industry is forced to redesign interconnects mid-cycle, the cost to scale will balloon, significantly impacting free cash flow for the entire hyperscaler cohort.
"Indium phosphide shortages compound existing grid/permitting friction in power-constrained regions, potentially doubling effective delays beyond supply timelines alone."
Gemini's R&D write-off thesis assumes forced mid-cycle redesigns, but that's speculative. Hyperscalers have 2–3 year roadmaps already locked; silicon photonics R&D runs parallel, not serial. The real risk Gemini flags—FCF compression—is valid only if indium phosphide delays *exceed* substitution timelines. That's a 2028+ problem, not immediate. More pressing: Grok's grid constraint angle is underexplored. Midwest data centers already face transmission bottlenecks; adding mineral scarcity creates a *dual* permitting/infrastructure wall that tariffs can't fix.
"Indium phosphide shortages could force a mid-cycle shift to silicon photonics, triggering sustained capex drag and FCF compression beyond simple substitution costs."
Gemini's R&D write-off thesis risks underestimating timing and costs: a forced mid-cycle shift to silicon photonics due to indium phosphide shortages could trigger a sustained capex drag and FCF compression across hyperscalers, not just a one-off substitution cost. If suppliers pivot slowly and integration hurdles mount, the 2026–2029 build-out may stall more than assumed, challenging the market's view that margins alone will absorb these risks.
The panel agrees that China's indium phosphide export curbs pose a near-term cost and delay risk to U.S. AI infrastructure, but not an existential threat. The main concern is the potential multi-year headwind for data center CapEx due to material scarcity and social engineering combined.
Accelerated domestic mining and onshore capacity development to reduce long-run dependency on foreign supplies (ChatGPT)
Forced mid-cycle redesigns of interconnects due to indium phosphide shortages, potentially leading to R&D write-offs and significant impacts on free cash flow for hyperscalers (Gemini, ChatGPT)