An Nvidia employee has been detained in Taiwan over allegations of forgery and breach of trust, in relation to the Supermicro smuggling scandal, that saw servers ostensibly sold to companies in Southeast Asia routed to China instead. Nvidia itself hasn't been accused of wrongdoing, and it published a statement calling smuggling a "nonstarter," saying that any GPUs sold through such a system would have no "service, support, or updates."

But that hasn't stopped Nvidia from taking its own measures to reduce its exposure to potential future smuggling efforts. Earlier this month, it created a form of "whitelist" for companies it sells to. It also investigated the firms it will continue to do business with, even sending staff members to customer data centers at the urging of the White House for verification.

Prosecutors have made it clear from the start that Supermicro isn't under investigation, merely its employees. The same is true of Nvidia. But as the AI frontier model race heats up and the White House floats banning Chinese models outright, Nvidia could face further restrictions on its hardware sales and greater scrutiny of its international actions.

Investigation escalation

The Supermicro smuggling scandal first came to light in March, when a trio of individuals were detained for deliberately mislabelling servers planned for sale to Southeast Asian countries. Instead, though, they sold them to China, getting around US export controls. The detentions included Supermicro co-founder, Yih-Shyan "Wally" Liaw, as well as a Supermicro sales manager in Taiwan, and a third-party broker who previously worked at Supermicro.

Where those detentions happened on U.S. soil, though, the investigations went international in May, when the Taiwan Keelung District Prosecutors' Office executed search warrants against three individuals it claimed were involved in illicit smuggling efforts. Although it was said to be independent of the U.S.-led investigation, it involved the same companies and was part of the same overall scheme designed to smuggle Nvidia hardware into China.

In Taiwanese law, selling GPUs to China — even the U.S.-restricted kind — isn't strictly a crime, but filing fraudulent paperwork and falsifying documentation absolutely is. That's why Taiwanese authorities have leaned on local fraud laws to tackle this increasingly international case.

Although the authorities were clear that Supermicro as a company wasn't being investigated, a number of high-level employees were. That continued in June when Taiwanese officials raided the Supermicro offices in Taiwan, as well as the homes of six individuals and three company sites, all said to be involved in the smuggling scheme.

The widening scope of the investigation ultimately pulled in workers from Supermicro distributor Albatron Technology and data center operator Chief Telecom. Taiwan has since said it is considering placing a criminal ban on all AI chip exports to China, locking down smuggling routes that have been actively exploited for several years.

But now the investigation is escalating up the supply chain and has now reached Nvidia itself. Although the company isn't under investigation, Nvidia's culling of potentially problematic suppliers and buyers might not do much if its own workers are facilitating the smuggling actions.

This is serious

The Nvidia employee in question has the surname Chang, but has remained otherwise unnamed. He was detained on suspicion of falsifying business documents, with authorities searching his home and workplace on July 24, marking the first time that Nvidia's premises have been investigated in this manner since the start of the smuggling scandal.

Prosecutors consider him strongly suspected of the charges, with a very real risk for attempted flight, destruction of evidence, and collusion with witnesses.

"Smuggling is a nonstarter," an Nvidia spokesperson told Tom's Hardware. "We primarily sell our products to well-known partners, including OEMs, who help us ensure that all sales comply with U.S. export control rules. Even relatively small exporters and shipments are subject to thorough review and scrutiny on both sides of the globe, and any diverted products would have no service, support, or updates."

Although authorities are clear that they are not investigating Nvidia as a company, an employee's involvement in the scheme will put a spotlight on Nvidia's actions and raise further questions about any additional involvement it or its employees may have had.

CEO Jensen Huang said in May that there was "no evidence of any AI chip diversion," but the situation has obviously changed since then. At the beginning of June, U.S. Senator Elizabeth Warren wrote to Nvidia general counsel Tim Ter, asking for evidence that supported Huang's claims.

Supply and demand

At the time of writing, there is a legitimate channel for Chinese firms to purchase Nvidia GPUs, but they're not the most cutting-edge Blackwell chips. There are older Nvidia GPUs granted licenses that are reviewed on a case-by-case basis, with the U.S. government taking a 25% revenue share cut of the sales. This reportedly adds up to just 75,000 units for 10 different Chinese companies - a relatively trivial amount of GPUs for Nvidia.

This is for the China-only, neutered Nvidia GPUs like H20 and H100s — not the cutting-edge GB200 and GB300 Blackwell-based stacks available to Western AI developers.

But this legal demand comes despite the lack of cutting-edge hardware options, the regulatory hoops that those involved need to jump through, and the Chinese government using carrots and sticks to encourage the use of domestic chip options.

That's because for certain tasks, Nvidia GPUs remain the best. For training, there's nothing that can compete with Nvidia's options. Chinese firms like Deepseek have tried previously, but they had to switch back to Nvidia when Chinese alternatives didn't measure up. Although some post-training fine-tuning is now possible on Chinese hardware, the Moonshot's headline-grabbing Kimi K3 was trained on potentially smuggled Nvidia Blackwell GPUs.

Considering the impact that Kimi K3 has had on the AI industry, it's hard not to imagine other Chinese AI developers looking to have their own "Deepseek moment" wouldn't search out access to Blackwell GPUs themselves.

The net may be closing on the Supermicro smuggling scheme, but the incentive is there for others to take its place, if they haven't already.

Police from Emporia, Kansas, arrested a teacher for applauding a speaker who was speaking out against a proposed zoning change to accommodate a planned data center in the area. According to 12 News, the police dragged 37-year-old Lux Claridge from the community meeting and charged him with disorderly conduct and interference with law enforcement. The commissioners had reportedly repeatedly warned against clapping and making other reactions while someone was speaking.

A video circulating on the internet shows Claridge clapping five times after a speaker closed their statement. We can hear in the clip below someone saying, “Ask him to leave,” followed by a second voice confirming, “Can I?” The voice then asked a police officer, “Chief? Will you ask-will you take the next person out that claps or anything, please? Thank you.”

The cops then approached Claridge, who said, “I have a right to speak.” The police were nonetheless insistent, which is when he said, “Drag me out.” Four officers then proceeded to cuff the teacher and hoisted him off to Lyon County Jail.

The data center’s critics warned against the project, saying that it could strain the local power supply and negatively impact the residents’ water quality. Their concerns are not without merit, as AI data centers have caused a massive 76% increase in electricity costs in the U.S.’s largest power region, while a Meta site is accused of muddying an entire town’s water supply.

Despite that, Emporia, located about 100 miles Southwest of Kansas City, approved the zoning changes needed for the Flint Hills Digital Campus project to move forward with the permitting process. This does not mean that the data center will immediately start construction, though, as it probably still needs to go through several more steps before it can break ground.

This isn’t the first time that a data center dissenter was hauled off to jail for reasonably pushing back against the said project. An Oklahoma farmer was arrested back in April for trespassing after going a few seconds over his allotted time and handing paperwork to the commissioners. Pushback against projects like these have been happening across the country, with several jurisdictions like Seattle and New York State enacting one-year moratoriums to study the potential effects of these infrastructure projects and how they could be mitigated.

President Donald Trump created the “ratepayer protection pledge” to force AI hyperscalers to “pay their own way” when it comes to electricity costs. He has even expanded this recently to include states and utility companies, with 23 governors and 187 firms signing up on the promise. Unfortunately, this is just a piece of paper and has no legal or regulatory power over the signees, with tech companies pushing back against a California bill that wants to turn this into law.

At the moment, Oregon is the only state that has passed and enacted a law that forces electricity consumers that used 20 MW or more to cover their fair share. Because of this, Portland General Electric, the state’s biggest power generator, has increased data center bills by 30% while cutting residential electricity costs by 1.3%.

Claridge is currently out on bail and is awaiting his hearing in September. "I'm glad to be out, but this is an inconvenience, really," Claridge said to local media. "It's not really deterring me from speaking out or, I guess, clapping."

SK hynix reported second-quarter revenue of 79.32 trillion won and operating profit of 60.54 trillion won on Wednesday, the latter up 557% year over year at a record 76% operating margin, and used the same Seoul earnings call to lift its 2026 capital spending guidance to the high 40 trillion won range as AI server demand keeps outrunning what the company can produce. Third-quarter DRAM bit shipments are guided up around 10% sequentially, following a quarter in which DRAM average selling prices rose roughly 30%, and NAND prices rose in the mid-50% range.

SK priced 177.9 million American depositary receipts at $149 each earlier this month, raising $26.51 billion in the largest share sale by a non-U.S. company on record, with the SEC filing earmarking proceeds for Korean manufacturing facilities and equipment, including EUV scanners.

Wednesday's capex number is roughly the same size and funds an accelerated mass production schedule at the M15X fab in Cheongju, the Yongin Phase 1 cleanroom that opens in early 2027, and the previously announced P&T7 advanced packaging plant and M17 NAND base, which SK hynix said will be built in phases according to customer demand. Cash and short-term investments hit 88 trillion won at the quarter's end, up 33.6 trillion won in three months, against interest-bearing debt of 18.6 trillion won and a debt-to-equity ratio of 7%.

CEO Kwak Noh-jung called 2027 the worst year of the shortage on the day of the Nasdaq listing and put the end of the crunch beyond 2030. Full-year DRAM demand is growing at a mid-20% rate by the company's own estimate, against bit shipments guided up around 10% next quarter. None of the capacity now being funded will produce wafers before 2027.

Operating profit landed below the 64.1 trillion won that brokerages surveyed by Yonhap Infomax had modeled. Executives attributed the softer blended DRAM ASP to product mix and to high-value shipments pushed into the second half, and said the gap should close as HBM4 and 1c-node conventional DRAM ramp up. HBM4 entered mass production during the quarter, and HBM4E samples have shipped, with volume production targeted for 2027.

Triggered by global selloffs, SK hynix closed down around 10% in Seoul on Wednesday, and Samsung Electronics fell 5%, with the KOSPI ending the session 6% lower and below 6,000 for the first time since April 14. The index touched 5,262 at one point, down almost 13%, taking its five-session decline to 17% and cutting a year-to-date gain that had reached 116% in June to 34%. Over the past month, SK hynix has lost 47% of its value and Samsung 37%. The KOSPI fell 10.84% on Tuesday and triggered a marketwide circuit breaker after SK hynix's American depositary receipts dropped below the $149 price at which they listed on Nasdaq on July 10.

Weaker shareholder-return expectations compounded the earnings miss, with SK hynix telling analysts only that additional returns remain under evaluation and would be disclosed within the year. Josh Gilbert, eToro's lead analyst for Asia-Pacific and the Middle East, told Bloomberg that "expectations had simply moved ahead of what even another record quarter could deliver."

CXMT closed its Shanghai debut up 466% on Monday after raising 57.92 billion yuan for DRAM wafer lines, and a report last week put China at low-volume production of domestic immersion DUV scanners running to around five units this year. TrendForce still has conventional DRAM contract prices rising 13% to 18% in the third quarter, with NAND up 10% to 15%.

The drone arms race is set to enter U.S. schools with Mithril Defense’s Campus Guardian Angel drones in at least nine schools across three states before the year is out. The Washington Post reports that the pilot program will be tested in Florida, Georgia, and Colorado schools with and without resource officers on campus. Securely siloed drones will be situated in key areas around campus, triggered by teachers via app or panic button, then fly to combat gun-toting attackers with a mix of strobes, sirens, pepper spray, and a 60 mph attacker-ramming capability.

If this is a successful initiative, it could go some way to lift the terrible specter of gun violence overshadowing U.S. students. The Washington Post notes that since Columbine in 1999, almost 400,000 students have experienced gun violence at school.

The Campus Guardian Angel drone rollout will be quite expensive. For example, the five schools in Georgia set to adopt drone defense will get $500,000 in backing to run the pilot program. However, politicians and legislators in that state appear to have passed the budget with ease.

This is how the Campus Guardian Angel drone system is designed to work:

• Drones are sited at secure locations around the campus,
• A teacher will trigger the drone deployment using an app or panic button,
• Remote pilots take control and distract, deter, and disable aggressors using a mix of strobes, sirens, pepper gel, and 60‑mph impacts,
• Drone system reaction speed is crucial.

Mithril Defense says that the Campus Guardian Angel drones can react and reach a shooter as quickly as 15 seconds. Justin Marston, Mithril’s founder and CEO, made a few other bold claims regarding his drone system’s reaction speed. For example, Marston said a drone like the Guardian Angel ‘might’ve saved lives in Uvalde.’ That’s a direct reference to the massacre at an elementary school in Uvalde, Texas, where 19 students and two teachers were killed in 2022.

The drone defense firm's CEO underlined that “the first 120 seconds are incredibly critical, because that’s when most of the shooting happens.” That implies that he thinks Campus Guardian Angel drones could be successfully deployed within that very narrow time window. But Uvalde was quite unusual, as responding law enforcement seemed paralyzed, waiting over an hour to enter the classroom.

These school-based drone systems are not without their critics. Some say that the funds may be better spent on prevention than cure. Even Mithril’s founder and CEO’s opinion seems to be that if these drones aren’t prompted into action within two minutes, they aren’t living up to their promise.

There is also the concern that drones could misidentify students or protection officers when controllers are under pressure. Others say that military-style drone systems aren’t appropriate for schools, and will cost a lot more than simple measures and routines regarding locked doors.

With the three states proceeding with pilot programs this year, we may see the true value and capabilities of the Campus Guardian Angel drones. If these drone-protected schools don’t suffer any terrible shooting incidents, then it may be claimed that the drones are at least a deterrent.

Keeping the upper hand in the AI arms race has become a vital goal for both the U.S. and China, and Nvidia's Blackwell AI chips are one of many flashpoints in that fight. The US government bars their sale to Chinese firms, while Chinese policies block their import as the country tries to spin up an advanced AI chip industry of its own.

But as we've discussed multiple times and then some more, Chinese AI firms are quite creative with workarounds for these restrictive policies. That's the case of Moonshot AI, which has reportedly made good use of Blackwell for training the recently released Kimi K3 frontier-level model, and is seemingly looking to obtain additional access in preparation for Kimi K4.

The Information says "people with knowledge of the matter" told it that Moonshot employed two Chinese firms that have Blackwell chips in their respective datacenters despite the bilateral restrictions we mentioned. Given that those chips are scarce enough right now even when obtained legitimately, it's unsurprising that neither firm had enough of them on hand to let Moonshot train K3. This reportedly forced Moonshot to figure out how to join multiple eight-chip Blackwell servers together and across datacenters in order to harness the necessary computing power.

The report also mentions "a researcher at a major Chinese tech firm who works on model training" as stating that Kimi K3 has "started a new round of arms race" in the country's AI industry. They further added that training frontier models is difficult or impossible with the promising but slowly developed homegrown chips. By that source's account, Chinese AI accelerators remain a generation or two behind Nvidia's current offerings and are reportedly several months in backorder.

For inference work, Moonshot reportedly relies on Nvidia's China-market HGX H20, a last-gen chip that isn't blocked by trade laws on either side of the Pacific. The firm recomends setups with at least 64 H20 GPUs for running Kimi K3. Those requirements, combined with that frontier model's desirability, meant that Moonshot quickly ran out of computing capacity to run K3 and currently has subscriptions on a waiting list. Given it's an open-weight model, and that its weights were released this week, many other inference providers are serving it, perhaps alleviating that bottleneck.

Meanwhile, White House Director Michael Kratsios claimed last week in a tweet that that Moonshot AI both "acquired GB300-equipped servers and has accessed GB300s in Thailand." While buying Blackwell chips is illegal, renting them is apparently fair game, at least until the proposed Remote Access Security Act takes effect. That law is designed to prevent the rental loophole by treating remote access as an export event. There's no telling exactly how the U.S. would enforce this law across other jurisdictions, though.

At any rate, the Department of Commerce is formally investigating if Chinese firms are accessing advanced U.S. chips like Blackwell GPUs, and that's likely to be an ongoing point of contention as the war for frontier model supremacy continues.

In China, it's an open secret that many of the country's high-level own or have access to Blackwell and other advanced chips, but despite all the trade restrictions and pushing the usage of local-made chips, the CCP has seemingly yet to crack down on said AI players. Some have theorized that the turning of this blind eye is intentional so Chinese firms like Moonshot can catch up to the likes of Anthropic and OpenAI.

Intel Foundry says that it has completed RAMP-C, the United States government program awarded to the company in 2021 to stand up a secure, domestic leading-edge chip ecosystem on its 18A process. The program funneled money to commercial and defense partners to run test chips on a design kit that wasn't finished yet, and its conclusion marks readiness for external customers who will pay full price to fabricate real products, including those who might use the Intel Secure Enclave manufacturing flow that RAMP-C helped to shape.

The announcement doesn't name any customers that Intel may have secured as a result of RAMP-C, although the program roster has been public for years and includes Nvidia, Microsoft, IBM, Qualcomm, Boeing, and Northrop Grumman. That may be down to the sensitive nature of any actual defense industrial base (DIB) products that are likely to be produced with the Secure Enclave defense-focused manufacturing flow for those products, which spans project stages from design to chip fabrication to advanced packaging.

Intel says RAMP-C was one of the programs that influenced Secure Enclave, and having a number of potential DIB customers run and test prototypes on 18A through likely generated valuable knowledge for Intel, the U.S. government, and its partners as those stakeholders work together to create a defense-ready domestic chip source.

Stu Pann, then SVP and GM of Intel Foundry Services, told Tom's Hardware back in February 2024 that the program's announced partners were IBM, Microsoft, and Nvidia, and that all three were running test chips paid for by RAMP-C. The funding let them "operate with immature PDKs, which normally they wouldn't do," Pann said, and covered their associated costs. Intel got PPAC data in return: how prospective customers rated 18A on power, performance, area, and cost.

Intel disclosed the roster in stages rather than all at once. Nvidia, Qualcomm, Microsoft, and IBM were named across the first two phases, Boeing and Northrop Grumman joined in July 2023, and Trusted Semiconductor Solutions and Reliable MicroSystems came in under a third phase that Intel says was awarded in April 2024.

Nvidia's participation runs back to the program's early phases, four years before it agreed to buy $5 billion of Intel common stock in September 2025, a deal that closed in December at $23.28 per share for more than 217.4 million shares. That agreement covers custom x86 CPUs and RTX SoCs, and carries no commitment to manufacture Nvidia silicon at Intel.

Secure Enclave, the follow-on program Intel references, is worth up to $3 billion and was finalized alongside the company's $7.86 billion CHIPS Act award in November 2024. Congress required that CHIPS money pay for it, which is why the commercial grant is smaller than the $8.5 billion originally proposed.

Intel Foundry booked $293 million in external revenue last quarter against $5.8 billion in total segment revenue and a $2.1 billion operating loss, according to the company's Q2 2026 financial results. Fortinet, named last week as the first publicly disclosed external foundry customer under CEO Lip-Bu Tan, is building its security processor on Intel 4 rather than 18A. On the same earnings call, Intel committed to 14A high-volume manufacturing in 2028.

OpenAI CEO Sam Altman recently declared on the Relentless podcast that artificial intelligence has entered the technological singularity, telling the show, "we are now, like, in the singularity," and that he'd been waiting for the moment his whole life. Two weeks ago, OpenAI’s own models broke out of a locked test environment and hacked Hugging Face's production servers.

OpenAI's account of the July 11 breach says GPT-5.6 Sol and an unreleased model were "hyperfocused" on the ExploitGym benchmark and went to "extreme lengths" to complete it. Rather than solve the exercises, they spent what OpenAI describes as a substantial amount of inference compute finding a route to the open Internet, exploited a zero-day in a package registry cache proxy, moved laterally through OpenAI's research network, and pulled the test solutions straight out of Hugging Face's production database. The company took ten days to tell Hugging Face who was responsible.

The definition of the singularity that Altman is invoking is set out by mathematician I. J. Good in 1965 and named by Vernor Vinge in 1993, rests on recursive self-improvement: a machine that designs a better successor, which designs a better one again, outstripping human intelligence.

Back in February, OpenAI told investors its inference expenses rose fourfold during 2025, dragging adjusted gross margin down to 33% from 40%. The same report put OpenAI's 2025 revenue at $13 billion against a target of roughly $600 billion in total compute spend through 2030, and Altman has separately committed to $1.4 trillion for 30 GW of capacity. The company has already walked back its first-party data center ambitions in favor of leasing.

Security firm Hacktron benchmarked GPT-5.6 Sol Ultra, Sol Medium, and Grok 4.5 on Chrome exploit development earlier this month, processing 2.096 billion tokens across the run, with one model finishing a complete exploit chain. An intelligence explosion should show capability per unit of compute climbing steeply, and OpenAI's own figures show the cost of a unit of capability going up.

Demis Hassabis, CEO of Google DeepMind, closed Google I/O in May by telling the audience they were standing in the foothills of the singularity. Altman's June 2025 essay, "The Gentle Singularity," had already placed humanity past the event horizon a year earlier, but neither Hassabis nor Altman named a threshold that would settle the question.

Aikido Security tested 13 models against 26 known CVEs this month and found GPT-5.6 topping the field at 23 of 26, or 88.5% recall. Moonshot's open-weight Kimi K3 matched that score at pass@3 for less money per run.

Google's parent company Alphabet recently reported negative free cash flow of $5.9 billion for the second quarter of 2026, the company's first cash-negative quarter since its 2004 IPO, after capital expenditures doubled year-over-year to a record $44.9 billion and exceeded the $39.1 billion its operations generated as it continues its rapid buildout of AI data centers, according to its earnings release.

CFO Anat Ashkenazi raised full-year capex guidance to between $195 billion and $205 billion, up from $180 billion to $190 billion, and disclosed that Google delivered TPU systems to customers' data centers for the first time, a shift from renting the chips exclusively through Google Cloud.

The quarterly deficit is small compared to the sums moving through the business, and the firm's trailing 12-month free cash flow remains positive at $53.3 billion. Back in February, Alphabet raised its guidance, but since then, spending has exceeded the cash the business generates due to its AI buildout, and Alphabet is covering the difference with borrowed money and new stock.

Servers first, buildings second

Approximately 60% of the quarter's technical infrastructure investment went into servers, with the remaining 40% split across data centers and networking equipment, Ashkenazi told analysts on the earnings call. That ratio inverts the usual assumption that hyperscaler capex is dominated by construction. Most of Alphabet's marginal dollar now buys compute, primarily its own TPU-based systems, rather than other forms of infrastructure. Depreciation of property and equipment rose to $7.1 billion in the quarter from $5.0 billion a year earlier, and Ashkenazi said infrastructure spending will keep pressuring the P&L through higher depreciation and energy costs.

"We're still in a supply-constrained environment," Ashkenazi said on the call, repeating a characterization the company has used for several consecutive quarters. Demand is running far enough ahead of Alphabet's own build schedule that the company is renting third-party capacity as a bridge while its data centers come online, an arrangement Ashkenazi said will create modest margin pressure for its Cloud segment in Q3. The construction pipeline behind the 40% includes a $40 billion, three-campus program in Texas through 2027 in November, representing the company's largest investment in any state, and a $1.5 billion expansion of its Jackson County, Alabama campus, announced in June.

TPU sales turn capex into inventory

Google began recognizing revenue from TPU system sales in the quarter, with Ashkenazi telling analysts the systems were "delivered to customer data centers for the first time in Q2" and that "the vast majority of the revenues from these agreements will be realized in 2027." According to Google's balance sheet, inventory stood at $10 billion on June 30, roughly four times the $2.4 billion recorded at the end of 2025. A meaningful slice of the quarter's cash outflow bought hardware that sits on the balance sheet today and will be sold to customers next year, bringing cash back in. Money spent on data centers doesn't return in the same manner; instead, it is written down over the years of use.

Anthropic is anchoring a great deal of Alphabet's external demand, with its October 2025 agreement giving the Claude developer access to up to one million TPUs and more than 1 GW of capacity coming online this year, and an April securities filing from Broadcom, Google's TPU co-designer, added roughly 3.5 GW of TPU capacity from 2027 while locking Broadcom into future TPU generations through 2031. Meta entered talks for multi-billion-dollar TPU deployments in its own data centers last November. The current flagship, the seventh-generation Ironwood TPU, carries 192GB of HBM3E per chip and scales to 9,216-chip pods that Google rates at 42.5 FP8 exaflops.

Every TPU Google manufactures serves four functions: training and serving Gemini, running Search and YouTube inference, renting to Cloud customers, and now shipping as sold hardware. No other hyperscaler's capex spend works that many jobs, and none of the others has a chip business generating third-party revenue at this stage.

A $98 billion liability

Alphabet issued Class A, Class C, and mandatory convertible preferred stock in June for net proceeds of $49.6 billion, earmarked in the release for "capital expenditures to scale AI infrastructure and global compute," and sold $20.3 billion of senior unsecured notes during the quarter. Long-term debt reached $98.2 billion on June 30, up from $46.5 billion at the end of 2025 and from roughly $16 billion a year before that, a run-up Ashkenazi acknowledged on the call. The February bond program alone raised more than $30 billion across multiple currencies, including a 100-year sterling tranche, it was reported at the time.

Together, the four largest hyperscalers plan a combined 2026 capex of around $725 billion, up 77% on 2025, and Alphabet's new range now tops the group alongside Amazon's roughly $200 billion. Meta raised its own 2026 forecast to $125 billion to $145 billion in April, citing component pricing and competition for land, power, and labor. Alphabet's headline Q2 net income of $112.1 billion overstates the reality somewhat, however, as $99.0 billion of other income came primarily from unrealized gains on equity securities, contributing $6.26 of the $9.11 in diluted EPS. Operating income, the cleaner measure, rose 30% to $40.8 billion.

Google Cloud grew 82% to $24.8 billion in the quarter with an operating margin of 35.6%, and backlog reached $514 billion, up more than $50 billion sequentially, with just over half expected to convert to revenue within 24 months. Those contracts are the collateral behind the spending, with the buildout chasing demand Alphabet has already booked rather than demand it hopes to find. Ashkenazi said free cash flow "will remain under pressure" and confirmed capex will rise significantly again in 2027, so the question the next few quarters will answer isn't whether Alphabet returns to positive territory in any given period, but whether operating cash flow, up 41% year over year in Q2, can keep growing faster than a spending that shows no sign of slowing down.

Taiwan's Keelung District Prosecutors' Office said on Tuesday it has detained a man surnamed Chang on suspicion of falsifying business documents, after investigators searched his home and his workplace on July 24 in connection with the AI chip smuggling case it opened in May. Bloomberg reports that Chang works for Nvidia and that the workplace search covered his desk at the company's Taipei office, which would make this the first known legal action against an Nvidia employee in a chip diversion case. Prosecutors said they consider him strongly suspected of the offenses and cited risks of flight, destruction of evidence, and collusion with witnesses, but haven’t accused Nvidia of any wrongdoing.

"Smuggling is a nonstarter," an Nvidia spokesperson told Tom's Hardware. "We primarily sell our products to well-known partners, including OEMs, who help us ensure that all sales comply with U.S. export control rules. Even relatively small exporters and shipments are subject to thorough review and scrutiny on both sides of the globe, and any diverted products would have no service, support, or updates."

Taiwan doesn't treat the unauthorized export of AI chips to China as a crime, which is why every detention in the case has so far concerned fraud accusations. The three people arrested in May were pursued over shipping declarations rather than the shipments, and the six summoned during June's raids on Super Micro and two supply-chain partners were questioned on the same basis.

Huawei and SMIC were among 601 entities added to the International Trade Administration's strategic high-tech commodities entity list in June last year, a designation that requires government approval before a Taiwanese company can ship to any of them. The list works off buyer names and carries no performance threshold, so a rack of accelerators sold to a company that isn't on it needs no approval.

It was reported last month that Taipei is weighing performance-threshold controls modeled on Washington's, and the Ministry of Economic Affairs confirmed consultations with the U.S. on bringing advanced chips under regulation without setting a timeline. Seven weeks on, prosecutors are still building the case out of the Criminal Code.

Senator Elizabeth Warren, ranking member of the Banking Committee, wrote to Nvidia general counsel Tim Teter and audit committee chair Brooke Seawell on June 1 asking what records support Jensen Huang's public claim that "there's no evidence of any AI chip diversion," and whether the committee had reviewed export compliance following March's indictment of Super Micro co-founder Yih-Shyan "Wally" Liaw over roughly $510 million in diverted servers. She set a June 18 deadline for answers.

A legal channel into China has existed since December, when the Bureau of Industry and Security began reviewing H200 export licenses case by case under a 25% revenue share, clearing around 10 Chinese buyers for up to 75,000 units each. Commerce Under Secretary Jeffrey Kessler told the House Foreign Affairs Committee on July 14 that shipments under those licenses remain trivial. Blackwell parts stay off the table entirely, and Chinese demand for them has pushed five-year-old A100 servers to $82,000 on the domestic gray market.

Having contributed to the growing shortage of memory and storage, AI companies seemingly have a new target in their sights: humanity's literary history. A recent investigative report from 404 Media reveals that these companies are reportedly purchasing millions of secondhand books through intermediaries to source high-quality training data for their AI models, avoiding public backlash.

AI relies on vast amounts of data to advance, but not just any data. It has to be high-quality data. The problem is that mediocre AI-generated content, commonly referred to as "AI slop," has proliferated across the Internet. This type of content contaminates the data pool and is counterproductive for AI to train on. As a result, leading AI companies have turned to human-authored sources for knowledge, specifically print sources that predate 2022 and are more likely to contain original, uncontaminated content.

There is precedent for AI companies turning to physical books for training AI. For instance, Anthropic, one of the leading AI companies involved in a lawsuit, reportedly invested millions of dollars in extracting information from countless printed books to build its Claude AI models and then destroying them. The company bought books from Better World Books. Although the court decision affirmed that using books for AI training falls under fair use in copyright law, Anthropic faced a staggering $1.5 billion fine for maintaining a repository of seven million pirated books that infringed the copyrights of authors and publishers. Similarly, a coalition of publishers recently filed a lawsuit against Google, accusing the tech giant of allegedly and illegally using millions of copyrighted books to develop its Gemini AI models.

ISBNdb, an online database that reportedly has over 111 million cataloged books, has been a long-favorite platform for booksellers, libraries, and distributors to sell books. With the explosion of the AI industry, ISBNdb has pivoted its business to offer specialized services to bulk-purchase books for AI companies. According to 404 Media, the orders range from 1,000 copies to as many as one million books in a single transaction.

One professional bookseller, who wanted to remain anonymous, purportedly spoke to 404 Media about the unprecedented surge in book sales, which began in April of this year. The seller previously moved around 20 books in a good week, but in recent months, weekly sales have skyrocketed to several hundred books. It represents a fivefold increase over the normal volume. Other booksellers on platforms such as Alibris and Biblio have reported similar spikes in bulk purchases.

While there is no concrete proof that ISBNdb or some other AI company is making the purchase, there are some red flags. Notably, the large-scale purchases only included books with an International Standard Book Number (ISBN), the unique 13-digit code used globally to identify books. There were no patterns in terms of subject, genre, or author. It also appeared that the purchasers disregarded the pricing for the books and snapped up titles at any cost, even if they were overpriced.

During the Anthropic lawsuit, Tom Harvey, who previously participated in the creation of Google Books before leading Anthropic's "Project Panama" digitalization project, confirmed that the AI firm hired several document scanning companies. Datamation Information Services, which offers high-volume, non-destructive, and destructive book scanning services, was one of them. The former method employs different tools, like overhead scanners, flatbed scanners, or V-shaped imaging systems. The latter method, on the other hand, would have personnel gut the books and feed the individual pages into a high-speed industrial scanner. Logically, AI companies opt for the destructive route since it is more efficient and lower-cost. The result is the destruction of millions of books.

Obviously, printed books represent a treasure trove of information for AI. However, many debate the ethics of removing books from circulation since it is uncertain whether AI companies filter the rare or even out-of-print books from the common titles during digitalization. The other major issue is that scanned books go directly into a private database to train AI, which the general public does not have access to. True, we will have smarter AI, but at the cost of the information not being available to future generations.

A coalition of over 30 tech industry leaders, including Nvidia, Microsoft, SpaceX, The Linux Foundation, Adobe, and Siemens has formed the “Open Secure AI Alliance” with the aim of building and distributing open source tools for AI safety and security, according to an official Nvidia blog post on Monday. The Nvidia-led coalition — comprising a mix of infrastructure, cloud computing, cybersecurity, and enterprise software leaders — will serve as a collaborative effort to develop open tools for identifying and patching AI vulnerabilities, sharing security frameworks, and establishing identity verification and audit standards across the AI software stack. Curiously, some of the biggest names in AI, including OpenAI, Anthropic, and Google, are absent from the list of members.

“The world needs both closed and open models. For cybersecurity, open models and open harnesses are essential because they democratize defensive capabilities, increase transparency for defenders, enable cyber defense while protecting data, and complement frontier closed models with customizable, localized controls. Open source enables massively distributed community-driven and self-controlled defense – with no single point of failure,” the announcement reads. Contributors across the alliance are currently building or offering various tools to create an open defense stack.

The initiative was directly galvanized by the OpenAI HuggingFace security incident earlier this month in which an autonomous OpenAI test agent slipped out of its sandbox and breached the AI startup Hugging Face. During the incident, safety guardrails on several frontier closed models prevented developers from performing critical forensic analysis. Hugging Face eventually had to use GLM-5.2 — an open-weight model from Beijing-based Z.ai — running the model on its own infrastructure to analyze more than 17,000 actions and contain the intrusion.

Based on the incident, the alliance contends that being unable to inspect, modify, or run a model locally — impossible in closed systems but doable with open systems — presents a fundamental weakness in relying exclusively on closed AI systems for cyber defense. The Open Secure AI Alliance therefore aims to give entities access to advanced open models, agent harnesses, and security tools that they can independently deploy and adapt, reducing dependence on any single provider while strengthening defenses across a multi-vendor AI ecosystem. “That is the mission of the Open Secure AI Alliance: to ensure defenders everywhere have open, frontier tools they can trust and control,” the post says.

Chinese models such as DeepSeek and the newly released Kimi K3 are open-weight, and are seeing growing adoption, including by U.S. companies, due to their open features. Meanwhile the Trump administration is reportedly gearing up to ban Chinese AI models over security concerns. The alliance acknowledges the potential risks of open source tools but argues that closed systems are not an outright solution. “Those risks are real, but they do not disappear in closed systems, and simply keeping weights closed does not prevent determined attackers from seeking or exploiting powerful AI,” the announcement reads. Unlike regulators who have voiced concerns over open-source technology, the alliance urges policymakers to treat open-weight models as defensive assets rather than liabilities.

It also argues that placing AI development solely in the hands of a few closed providers creates dangerous single points of failure. “The right response is not to deny defenders access to capable open systems. It is to pair openness with strong safeguards, clear rules against malicious misuse, rigorous evaluation and rapid remediation. Defenders need both frontier closed models and frontier open models, working together, so they can choose the right system for the job and ensure that transparency, adaptation and sovereign control are available wherever security demands them,” the alliance contends.

According to the announcement, “The Open Secure AI Alliance — building on the leadership of the Linux Foundation’s Akrites initiative and OpenSSF community work — will work to remediate and disclose vulnerabilities using open technologies”. Founding members include NVIDIA, Dell Technologies, Synopsys, Microsoft, IBM, Red Hat, CrowdStrike, Palo Alto Networks, Cloudflare, Hugging Face, Databricks, SpaceXAI, and The Linux Foundation. Conspicuously absent from the alliance are OpenAI, Google, and Anthropic, companies behind proprietary, "closed" AI models.

Well, the artificially intelligent cat is out of the bag. After publishing a blog post and API documentation for the minty-fresh Kimi K3, Chinese outfit Moonshot AI delivered on its promise to release the model's weights for free, meaning that most anyone with a contemporary rack of AI GPUs can run it and charge for it, with few restrictions.

This is quite the shot across the bow of the big AI players, namely but not only Anthropic and OpenAI. Those companies' latest models are Claude Fable and GPT-5.6 Sol, respectively, and it happens that Kimi K3's capabilities outright beat previous generations of Claude and GPT in Moonshot's benchmarks, and closely trail Fable and Sol— all while seemingly being around 2-3x cheaper to run, up to 10x if a particular query lands in the cache. Moonshot's technical write-up seemingly backs up the benchmarks published last week, as the company reveals which exact software was used for testing.

For its inference cost comparisons, Moonshot says that its costs "are measured internally" versus the publicly available token pricing for other companies, but the figures are quite impressive. For input, Moonshot charges $3 per million tokens for Kimi K3. Meanwhile, Fable costs $10/1M, while Sol goes for $5/1M. That figure is standard non-cached input and is already pretty good-looking, but Kimi K3's caching structure seemingly has a 90% hit ratio for coding tasks, turning those $3 into $0.30/1M if your use case hits the cache a lot. The story is pretty similar for output tokens.

One of the likely reasons why Kimi K3 is so efficient is that it uses a mix of MXFP4 for weights and MXFP8 for input activation, both data types with relatively low precision and thus amenable to running on far less VRAM. Out of Kimi's 2.8 trillion parameters, only 104.2 billion are activated at a time, too.

Interestingly, Moonshot's write-up only mentions Nvidia's H20 being used for running Kimi for some coding tests, a fairly low-end chip by today's standards. That GPU doesn't have native support for MX floating-point types, unlike the export-controlled Blackwell B-series chips.

In turn, this can mean that Kimi K3's optimizations make it particularly amenable to run on lower-end hardware, but it's an equally reasonable guess that running it on something like Nvidia Blackwell or other MXFP-native silicon could make it even more cost-effective than in the presented benchmarks. We'll have to wait for more official figures to confirm this speculation.

Additionally, Kimi K3 doesn't use a conventional ever-expanding key-value (KV) store, instead relying on a fixed-size state handler called Kimi Delta Attention, again theoretically saving both on VRAM and execution time. Its mixture-of-experts (MoE) is particularly sparse with only 16 activated at each time out of 896, further contributing to inference cost reductions. Broadly speaking, Moonshot went for optimization at every layer of inference to avoid unnecessary overhead and bring inference cost down.

This is could be bad news for OpenAI and Anthropic, given that most anyone with decent AI GPUs can now become their direct competitor, and the fact that Kimi K3 is open-weight also gives off the impression that "free" software is nearly as good, and far cheaper to run, than its proprietary competitors. It's worth noting that open-weight does not mean open-source; the training process and dataset are still Moonshot's special secret sauce.

A state-backed company in Shanghai has begun mass-producing immersion deep ultraviolet lithography machines and is due to deliver the first units this year to SMIC, Hua Hong Semiconductor, and memory maker ChangXin Memory Technologies, according to The Information, citing two people familiar with the program. Output targets around five machines in 2026 and roughly 20 in 2027, and all three named recipients sit on the list of Chinese firms that a bill now moving through Congress would cut off from ASML sales and servicing by statute.

The Information didn't name the manufacturer, but its sources described the operation as having pulled DUV development teams from several Chinese companies, one of them the state-backed startup Shanghai Yuliangsheng Technology. SMIC has been testing a Yuliangsheng immersion tool since September 2025. Most components in the new systems are domestic, though some critical parts still come from Japan, and delays at local suppliers have held back output this year.

U.S. House Resolution 8170 designates SMIC, Hua Hong, CXMT, Huawei, and YMTC as restricted entities in law, and three of those five are the named first customers for the domestic scanner. The MATCH Act, introduced in April, was reported out of the House Foreign Affairs Committee on April 22 and has a Senate companion filed as S. 4281. Its immersion DUV provisions cover servicing and technical assistance, not just new exports, which extends its scope to installed tools already operating in Chinese fabs, fabs which have spent the past two years stretching that installed fleet through secondary-channel upgrades.

ASML expects to ship about 130 immersion systems in 2026, matching 2025, CFO Roger Dassen told analysts during the company's July earnings call. Dassen added that ASML intends "to increase capacity by 30% in 2027" for immersion, and is investigating another 30% for 2028. China accounts for around 20% of ASML's net sales this year, down from 33% in 2025, driven mainly by mainstream logic demand.

Immersion DUV prints 28nm-class features in a single exposure and reaches 7nm through multipatterning, at a cost in overlay errors and yield. ASML CEO Christophe Fouquet told the same call that rising DRAM litho intensity partly reflects customers replacing multipatterning with cheaper single-exposure EUV.

Independent analysis from the AI Futures Project in June put commercial-scale Chinese immersion DUV in the mid-2030s, with ASML holding 98.7% of the immersion market. Qualifying the new machines for production lines could take many months, and they trail ASML's tools on performance and build quality. China's domestic EUV effort, which Reuters first reported as a working prototype in December, remains years away.

OpenAI is in advanced talks to lease SB Energy's 10 GW data center campus in Piketon, Ohio, with Nvidia in discussions to guarantee roughly $250 billion of the financing behind it, the Wall Street Journal reported on Sunday, citing unnamed people familiar with the matter. The site would be OpenAI's first as a tenant rather than a customer of Microsoft, Amazon, or Oracle, and Nvidia is separately discussing financing the accelerators going inside, which could run to another $350 billion. Terms haven't been settled, and the arrangement could still collapse.

OpenAI has no investment-grade credit rating, and Nvidia's involvement would let SB Energy raise debt against Nvidia's balance sheet instead of its tenant's. Nvidia has already put $30 billion into OpenAI, which has raised its projected compute spending to around $750 billion through 2030, up from roughly $600 billion earlier this year, according to the Journal. Commerce Secretary Howard Lutnick controls allocation of the site's power, and Anthropic, Microsoft, and Google have all spoken to him about it in recent weeks.

Nvidia's Q1 FY2027 10-Q caps maximum gross exposure across all of its partner facility lease guarantees at $3.5 billion, shrinking as partners pay their lessors, with $712 million sitting in escrow against it. Nvidia took the guarantees in exchange for warrants and carries them as credit derivatives, describing their fair value as immaterial.

The first one, disclosed in the third quarter of fiscal 2026, was capped at $860 million with $470 million of escrow behind it. The partner separately contracted to sell the data center cloud capacity, and Nvidia retained the option to assume the lease for internal use or sublease it if the escrow and that contract came up short. Neither remedy has an obvious equivalent at a 10 GW campus on federal land.

Nvidia held $62.6 billion in cash, cash equivalents, and marketable securities when fiscal 2026 closed on January 25, against full-year revenue of $215.9 billion and net income of $117 billion. A $250 billion guarantee works out at roughly 71 times the guarantee book Nvidia has disclosed, more than a year of revenue, and about four times its cash.

SB Energy broke ground at the former Portsmouth Gaseous Diffusion Plant on March 20 alongside Energy Secretary Chris Wright, Lutnick, and SoftBank chairman Masayoshi Son. The site enriched uranium for the U.S. weapons program from 1954 until 2001 and is still being decontaminated. The Department of Energy had listed it among 16 federal sites opened to data center construction, and SB Energy leases the land rather than owning it. Powering the campus takes 9.2 GW of new natural gas generation plus $4.2 billion of transmission work with AEP Ohio, funded by $33.3 billion Japan committed under its trade agreement with the U.S. The first phase, roughly 800 MW, is expected in 2028.

OpenAI gave up on building its own data centers last year in favor of leasing capacity, and SoftBank carries more than $130 billion of debt while funding buildouts in Ohio, France, and elsewhere.

When we talk about running local AI these days, the conversation usually either revolves around mini-PCs like the RTX Spark or drifts into wistful thinking about home servers and ludicrously expensive professional GPUs. Well, I reckon the most impressive AI hardware trick in a good while just happened on a piece of silicon that costs less than a decent burger. Last week, a Ukrainian developer named Slava S, who simply goes by 'slvDev' on GitHub, dropped a project called ESP32-AI. It's exactly what you think: he got a 28.9-million-parameter language model running locally, entirely on-device, on an ESP32-S3 microcontroller.

If you haven't read any of our previous coverage of this tiny chip, ESP32-S3 boards offer about the best bang for buck in the whole computing world. You can snag one online with a protective case for under $20 here in the States, and bare boards are readily available for under $10 around most of the world. As you'd expect from a chip so cheap, it's not powerful. On this variant, the S3, you get exactly 512KB of SRAM, 8MB of PSRAM, and 16MB of flash memory, which is not very much memory at all. So how exactly do you cram a nearly 30-million parameter model onto a chip with less primary storage than a single raw photo from your smartphone?

Usually, to run an LLM, the entire model has to sit in your system's fast memory because the processor needs to constantly do math against every parameter to generate the next word. If you try to run a 29M parameter model normally on an ESP32, you run out of fast RAM instantly. The previous record for a chip like this was around 260,000 parameters by one Mr. Dave Bennett, as pointed out by Slava himself on X.

Our clever hacker got around this bottleneck by borrowing a brilliant architectural trick from Google's Gemma called Per-Layer Embeddings. He quantized the model down to 4-bit (making the total file size just 14.9 MB) and changed where the data lives; instead of trying to stuff the whole thing into the tiny 512KB SRAM or the only slightly-less-tiny 8MB PSRAM, he dumped the 25-million-parameter embedding table into the relatively-slow 16MB Flash memory. Because this specific model architecture only needs to pull a few rows from this table per token, the inherent slowness of the Flash memory doesn't choke the processor, and so the 512KB of fast SRAM is kept clear for just the "thinking core", the actual reasoning weights.

Now, let's pump the brakes for a second, because I know someone out there is already wondering if they can replace their server with an $8 chip. The model he used was trained on the TinyStories dataset, and it's really more of a Small Language Model (SLM), or honestly, a "micro LM." Due to the way it was created, it's only capable of writing short, simple, fictional stories. It will not answer questions, it will not follow instructions, it won't write your Python code, and it possesses exactly zero factual knowledge about the real world.

Focusing on that limitation completely misses the magic of what's happening here in this proof-of-concept, though. The achievement is fitting a structurally quite large model onto a computer with practically no resources. It proves that with clever architecture, you can run genuine neural networks on dirt-cheap embedded hardware, and there are useful applications for a model this size. Slava imagines the idea of a coffee machine that actually knows about coffee: every bean, grind, ratio, water temperature, all offline, no app required.

Truthfully, when we're talking about "AI", it all comes down to what you are trying to accomplish. To put it plainly, asking how much hardware you need for local AI without specifying the workload is like asking what vehicle you need without saying what the goal is. A bicycle, a sedan, a pickup truck, a semi-trailer, and a train all "get you from A to B," but they're built for radically different jobs. AI is the exact same way; it's what you're doing with it that determines how much hardware you need.

To illustrate the point, last year another developer published a project cramming an AI language model (a bigram name generator) into the original Dragon Warrior and Final Fantasy games on the NES. Yes, the Nintendo Entertainment System. Developer Emanuele Rodolà managed to fit the entire model weight table (729 bytes) and the inference code (~140 bytes of hand-written assembly) into the original game ROM to generate new character names on the fly. That's real AI, running on a MOS 6502 processor, a piece of silicon that dates back to 1975.

The ultimate takeaway from slvDev's project is that Per-Layer Embeddings scale far further down than most people would have imagined, and it lends credence to the recent enthusiasm surrounding High-Bandwidth Flash as a tiered storage medium for AI servers. That's exciting not because it means an ESP32 will replace your desktop GPU, but because it suggests the same architectural ideas could make AI dramatically more practical across the entire spectrum of hardware, from tiny embedded devices all the way up to datacenter accelerators.

Residents of Dowagiac, Michigan, just filed a lawsuit against a data center that allegedly generates a high-pitched whining sound — and has done so 24/7 for the past two years.

The data center, which is owned by Alliance Cloud Services LLC, a subsidiary of Hyperscale Data, used to be an industrial building that sat behind a row of pine trees across the street from the most affected residents, according to ABC-affiliateWXYZ. The site started development in 2018 and was eventually turned into a cryptocurrency mining center in 2021. But in 2024, something changed drastically when the building started emitting noise pollution around the clock. Residents say it sounds like a vacuum cleaner in their living room, and the company has offered to buy up homes from unhappy residents.

“It sounds like someone set up a vacuum, like in your living room. And the vacuum is just... that thing needs to be cleaned ... the filter is clogged up, so it’s a high-pitch whining. And they just left it on and walked out," said Lindy Valenzuela, one of the residents living across from the data center. Billy Finn, who also lived nearby, added, “You’ve seen movies and stuff where they have somebody in a cell torturing them with sound. And that’s basically what it is."

Dowagiac has recently instituted an industrial noise ordinance, with a daytime limit of 65dB during the day and 55dB at night, and it has fined the data center for violations. However, Hyperscale Data is challenging the city’s readings and methodology. The company said that it’s planning to expand its operations in the footprint in the area, but the city said that it hasn’t received any permit applications.

Hyperscale Data CEO William Horne told the residents directly in a special council meeting that the site is pivoting away from cryptocurrency operations towards AI computing and advanced robotics, and that it’s spending $100 million to achieve this. It has also bought acres of adjacent properties for use as a natural buffer to reduce the noise that affects residents in the future, alongside other efforts that will reduce sound levels. "If they still aren't happy, and feel that their home isn't enjoyable, then we'll buy their property from them,” Horne said.

Still, this statement did not sit well with the affected residents. One of them said that their family has been living in the area for close to a hundred years, with their friends and support system, while another has a 17-month-old baby and another on the way, meaning moving for them is going to be difficult, if not impossible. Because of this, the residents asked the CEO why they didn’t act on the complaints as soon as they started and accused the company of not being a good neighbor.

This isn’t the first noise pollution lawsuit that a data center is facing in the country. A Microsoft data center is facing a similar class-action lawsuit in Wisconsin from residents who live within 1.5 miles of the facility. One non-profit organization also said that inaudible vibrations, called infrasound, that these industrial sites emit can be heard and felt for hundreds of feet in surrounding areas and could potentially have negative health effects on anyone who can feel it.

ChangXin Memory Technologies closed its first day on Shanghai's STAR Market at 49 yuan on Monday, up roughly 466% from an 8.66 yuan offer price, giving China's only volume DRAM maker a market capitalization of about 3.3 trillion yuan ($487 billion) and the top spot on the mainland market ahead of Industrial and Commercial Bank of China. The company raised 57.92 billion yuan ($8.6 billion) in Asia's largest IPO of 2026, and its prospectus assigns the bulk of the named project spending to wafer lines and process upgrades for DRAM it already produces, with nothing earmarked for high-bandwidth memory.

The prospectus splits 29.5 billion yuan across three projects: 13 billion yuan for DRAM technology upgrades, 9 billion yuan for next-generation DRAM research, and 7.5 billion yuan for memory wafer manufacturing line upgrades. The filing contains no dedicated HBM project and no disclosed funding commitment to a near-term HBM expansion, and CXMT hasn't broken down where the remaining roughly 28 billion yuan goes beyond describing it as working capital.

Conventional DRAM yields more than three times the bits per wafer that HBM does, and SemiAnalysis models CXMT's 8-high HBM3 yield at around 25%. Its cost per bit on DDR5 runs more than 30% above Samsung, SK hynix, and Micron. CXMT will add around 85,000 wafer starts per month of DRAM capacity this year, against 60,000 at SK hynix, 30,000 at Micron, and 15,000 at Samsung, per SemiAnalysis estimates.

That puts the company on course for roughly 350,000 wafer starts per month by the end of 2026, within 25,000 of Micron's total on Citrini Research's model, before a Shanghai fab two to three times the size of its Hefei headquarters reaches volume production in 2027.

It has been reported that output is already booked through the end of 2027, with DigiTimes having cited supply chain sources and Dell, HP, Lenovo, and Apple ahead of smaller buyers in the queue. CXMT signed a five-year server DRAM agreement worth more than $7 billion with ByteDance this month, and a $3 billion deal with Tencent in June, and server products grew from 8.4% of its revenue in 2024 to 26.5% last year.

Nomura opened coverage with a buy rating and a 116 yuan target, 1,239% above the IPO price, on an assumption that CXMT's share of global DRAM output climbs from about 10% now to 18% by the end of 2028. Morningstar puts fair value at 14.90 yuan, under a third of Monday's close, citing the company's lack of access to EUV lithography as the constraint on further conventional DRAM scaling. Nomura's downside case, built around potential equipment and materials embargoes, cuts 2027 to 2028 net profit by 30% to 33%.

Only 6.73% of CXMT's enlarged share capital was tradable at listing, and the lock-up expires on January 27, 2027. Buyers of the resulting modules aren't getting a discount, either. Retail DDR5 kits using CXMT dies track big three pricing, and early testing has shown the dies resist voltage scaling and overclock poorly next to SK hynix parts.

Imperial Valley Computer Manufacturing has filed a lawsuit in a bid to gain access to Colorado River water, 287 million gallons of which it says it needs to cool a 330-megawatt data center, which would be the largest in the state. Despite only representing a fraction of the region's water supply, the buildout of the data center may affect the local farming and adjacent industries and terminate hundreds, if not thousands, of positions, reports Business Insider.

After two cities in the region denied the California-based AI data center recycled wastewater for cooling, it filed a lawsuit demanding to get water from the Colorado River for cooling. The 330-megawatt facility was not only designed to be the biggest AI data center in California, but it specifically committed not to use water from the Colorado River because it was promised wastewater. But now the owner of the data center is essentially asking to redirect water supply from agriculture to the facility.

Imperial Valley Computer Manufacturing — the owner of the 330 MW AI data center — is requesting access to approximately 287 million gallons of water per year after two cities — El Centro and Imperial — declined to supply reclaimed wastewater for cooling. The Imperial Irrigation District (IID), which distributes Colorado River water throughout Imperial Valley, also denied the company's request. The Colorado River supplies water to roughly 40 million people across seven western states and serves as the valley's sole freshwater source for roughly 180,000 people. Agriculture consumes about 80% of California's allocation from the river, while roughly 95–97% of the water IID delivers goes to agriculture.

The data center is seeking roughly 287 million gallons per year (about 750,000 gallons per day, or ~880 acre-feet per year), whereas the Imperial Irrigation District (IID) holds rights to approximately 3.1 million acre-feet of Colorado River water annually, which means that the data center demands only a small fraction — 0.028% — of IID's total water supply.

Sebastian Rucci, a Huntington Beach attorney who leads the project, claims that the facility's water consumption would be comparable to that of a 160-acre farm and will require no additional Colorado River allocation. In fact, he states that the facility would not increase pressure on the river because the company intends to purchase nearby farmland together with its associated water allocations.

Under the proposal, irrigation on those properties would cease, thus transferring the existing water quotas to be redirected to the data center cooling, at the expense of local farming output and associated jobs. "There's a lot of resistance in any agricultural community to 'buy and dry' because that's jobs," a senior fellow at the Pacific Institute focused on Colorado River Basin water use told the outlet. According to them, local resistance to the plan is less about the amount of water, and more about buying up farmland and reallocating it for industrial use.

The approach, of course, differs from the earlier plan that intended to avoid using Colorado River water altogether. However, after the data center was denied wastewater from two cities, it does not have a choice if it wants to go ahead with the buildout.

Rucci reportedly indicated that the project would provide substantial economic benefits for the local community, including 1,688 construction jobs, more than 100 permanent positions, and an estimated $2.95 billion in economic impact over 30 years. For a region where unemployment stood at approximately 17% in May, the economic diversification is essential. However, the big question is whether 100 permanent roles could offset the lost positions in the farming industry and industries tied to agriculture.

Water policy specialists interviewed by Business Insider said that the debate extends beyond the project's annual consumption. Instead, they questioned whether converting irrigated farmland into industrial use is an appropriate long-term direction for the region, which has historically depended on farming. The experts also warned that although landowners could benefit from selling land or water rights, surrounding rural communities may lose employment and business activity adjacent to agriculture, which includes equipment suppliers, repair shops, and sellers of fertilizers. Another factor mentioned by the experts was the U.S. reliance on farms around Imperial, California, and Yuma, Arizona, as they were the main suppliers of certain agricultural products in winter.

Physicists at the Universities of Konstanz and Stuttgart have run chaotic-signal forecasting and anomaly detection on 400 microscopic particles orbiting in a drop of liquid, in work published in Communications AI & Computing. The array predicted a chaotic Mackey-Glass series and picked out anomalies that leave a signal's mean, variance, and short-time autocorrelation untouched, scoring an F1 of 0.90 on that harder task. It also came in roughly 10 times less accurate than memristor-based reservoirs, a gap the paper admits candidly.

Each oscillator is a silica sphere of 3μm radius, capped on one side with 80nm of carbon and suspended in a water-lutidine mixture held at 28°C. A 532nm laser heats the cap and drives the particle toward an assigned target point, but the delay between imaging a particle and repositioning the beam means it overshoots and settles into a small orbit instead. Flow fields in the liquid couple neighboring orbits, and data enters the system as displacements of the target points.

Lattice spacing sets coupling strength, since hydrodynamic forces fall off with distance, and a damping threshold sets how far each particle swings. Both are adjustable while the experiment runs, with a forecasting error that varies by more than a factor of three across that parameter space. Accuracy also held up when the input reached only 20% of the oscillators, and when individual particles stopped responding to the laser or clumped together.

The colloidal array reached a normalized root-mean-squared error of about 0.1 on the one-step Mackey-Glass prediction. Memristor devices now reach 0.01 or better on the same benchmark, the paper notes, adding that those results depend on time-multiplexing and follow nearly a decade of concentrated work.

The authors write that their reservoir doesn't outperform established physical implementations. An arXiv preprint from January, however, framed it differently, arguing that avoiding time-multiplexing set the platform apart from nearly all existing physical reservoirs, photonic, memristive, and spintronic ones included.

Running the reservoir takes a 532nm laser, a two-axis acousto-optical deflector scanning at 100 kHz, real-time microscopy with particle tracking, a temperature-controlled quartz cell, and a conventional computer for the 1,000 Gaussian kernels and ridge regression that produce the output. No energy stats appear anywhere in the paper, despite energy efficiency being the stated motivation, and the authors concede that the laser-driven setup might not be practically applicable and instead point toward simpler actuation schemes, such as electrode-driven colloids.

Clemens Bechinger, professor of soft condensed matter at the University of Konstanz, said in the university's announcement that the dynamics don't need to be fully understood, only to respond reliably, at which point "its physics can be directly harnessed for computation."

A separate team synchronized 105,000 nano-oscillators in 45 nanoseconds this month on a platform projected to run at tens of gigahertz.

MRI machines are life-saving medical devices that can let doctors and radiologists diagnose various critical conditions, but they’re also insanely expensive. Brand-new models start at $1.1 million and could go as high as $3 million per unit or more. The Open Source Imaging Initiative recognized this limitation and has been working on the open-source OSI2 ONE MRI scanner, which had already been replicated multiple times globally. However, this portable device, which has a 3D-printed core, has a limited field strength of just 50mT (compared to the 1.5T to 3T used by full-sized units). This gave them lower spatial resolution and lower signal-to-noise ratio, but tech analyst Brian Roemmele said on X that AI can overcome this and make it usable for medical diagnoses.

“Low-field MRI has historically been limited by lower signal-to-noise and greater field inhomogeneity. That is exactly the regime where modern AI thrives,” Roemmele wrote on the social media platform. “Image reconstruction becomes dramatically better when deep networks trained on high-field data or physics-informed models denoise, correct for inhomogeneity, and push resolution beyond the raw acquisition limits. Real-time sequence adaptation can adjust gradients and RF pulses on the fly as the AI monitors signal quality.”

Note that this isn’t just a general run-of-the-mill AI that everyone uses but a specially trained model on high-field MRI (1.5T to 8T) data or using the actual physics of the MRI machine so that it can create a more accurate picture. Scientists have already been using this technique for years, with some researchers training an AI model on 1.6 million brain scans to make it more accurate in detecting dementia. If an institution does not have access to anonymized patient data used to train the specialized AI, it can rely on synthetic data generation because of the open-source nature of the OSI2 ONE MRI scanner. Since all the information about the machine is publicly available, researchers could use this instead to build a physics model that the AI model can use.

Some people commented, saying that this won’t work in the highly regulated medical environments usually found in first-world countries. Nevertheless, Roemmele said, “No one can stop us from building in garages.” It also seems to be targeted for regions that have low access to technologies like these or do not have the financial capacity to purchase and maintain a full-sized device (even refurbished MRI machine units start at $100,000, and you also have to spend more to set up the specialized room that will house it).

While a portable MRI scanner like the OSI2 ONE will never have the resolution of the expensive, full-sized machines, it’s arguably better to have something that doctors can use for diagnosis without costing millions of dollars if the specialized AI model turns out to be effective and accurate. With that, even less wealthy hospitals and clinics could have access to this imaging device and save more lives. It also shows how the medical industry and even patients use AI to save on costs.