Posts Tagged ‘SK Hynix’

CXMT: Memory Giant, Or Smoke And Mirrors?

Tuesday, July 21st, 2026

This is a very different post than the one I set out to write.

When I talked about the current memory shortage, I didn’t really cover Chinese memory fabs, as I lumped them in with “older fabs” churning out older types of memory that weren’t relevant to the current memory crunch. However, after I posted, I came across news for ChangXin Memory Technologies (CXMT). According to this list of semiconductor fabs (which I’ve used as a starting point for many semiconductor posts), their Fab 1 in Hefei, China started running DRAM on a 19nm process node in 2019. That’s unsuitable to fab the high bandwidth memory (HBM) demanded by the AI buildout that’s currently making money for Samsung, SK Hynix and Micron hand-over-fist. But it is already churning out DDR4 (and low-power DDR4 for the mobile market), and just last year started fabbing DDR5. Naturally, being a Chinese company, they probably stole their technology from Samsung and SK Hynix.

CXMT just had a public offering on the Shanghai stock exchange that went very well.

ChangXin Memory Technologies (CXMT), mainland China’s only mass producer of DRAM—

Technically untrue. SK Hynix’s C2F fab in Wuxi, China turns out DRAM, but I included it in SK Hynix’s global output.

—officially launched its initial public offering on the STAR Market, raising approximately CNY 57.9 billion (roughly $8.4 billion) at CNY 8.66 per share, setting a new record for the board. Fueled by surging AI server demand, the company turned profitable for the first time in 2025, with net profit for the first half of 2026 estimated between CNY 50 billion and CNY 57 billion. Supply chain sources indicate that major PC manufacturers including Dell, HP, and Lenovo have already locked in CXMT’s DRAM production capacity through the end of 2027.

I haven’t been able to verify this tech name dropping, but I haven’t tried very hard, because I was chasing down more fundamental questions.

Meanwhile, although Apple is evaluating the integration of CXMT’s LPDDR5X chips, a Bank of America report points to three major constraints—geopolitical risks, technical specifications, and patent issues—making large-scale, substantial procurement unlikely in the near term. CXMT’s listing also provides a glimpse into China’s memory industry landscape, where breakthroughs are being made across DRAM, NAND Flash, and NOR Flash, though a significant gap remains compared to the world’s top three manufacturers.

I’ve seen the “Apple testing CXMT chips” in more reputable sources, so let’s assume it’s true for now.

$8.4 billion is a big chunk of change, but in the semiconductor world that will buy you…one 14nm fab, with a bit of change left over.

But according to some breathless analysis in the sort of farflung web outlets that don’t make up part of your regular feed:

A new report claims that CXMT will come exceptionally close to Micron’s DRAM wafer production capacity by the end of this year. In the coming years, China could become the world’s second-largest producer of DRAM globally, fundamentally altering the global DRAM market.

It has been estimated that CXMT will end 2026 with 350,000 wafer starts per month (WSPM). That’s a mere 25,000 fewer wafers than Micron. China has reportedly forced CXMT to share its DRAM technology with other Chinese semiconductor companies, including JHIC, Swaysure, and XMC (a YMTC subsidiary).

Obviously, you’re not getting 350,000 wafer starts per month from one fab. That piece tells us that CXMT has three fabs, two in Hefei and one in Beijing, with another in Shanghai due to come online later this year.

However, the source they cite is…a random guy on Twitter with 754 followers. I’d like a more solid source than that.

Could CXMT actually pay for two more modern-but-not-cutting edge fabs? Maybe. The city of Hefei (and possibly other government and CCP-connected entities) has been subsidizing them.

The fact that the additional fabs aren’t on the Wikipedia list isn’t dispositive. The Source of All Vaguely Accurate Knowledge relies on volunteer labor, and overwhelmingly English-speaking labor at that.

So let’s go to Google maps and take a look at the aerial view of them.

And here’s where the real trouble starts.

“ChangXin Memory Technologies Beijing” brings up nothing.

OK, let’s ask Google AI: “Where is ChangXin Memory Technologies semiconductor fab in Beijing?” Google’s answer:

ChangXin Memory Technologies (CXMT) operates its Beijing semiconductor fabrication plant (known as Fab C1/C2) within the Beijing Economic-Technological Development Area (BDA), also commonly referred to as the Yizhuang Development Zone in the southern suburbs of Beijing. Production at this specific memory/DRAM site began in 2022.

Hmmm. Let’s see if we can narrow that down: “What is the street address of ChangXin Memory Technologies semiconductor fab in Beijing?” Google’s answer:

The primary location for this facility is: Building 52, No. 2 Jingyuan North Street
Beijing Economic-Technological Development Zone
Beijing, China

Great! Let’s pop that into Google Maps.

And here we run into multiple problems:

Problem 1: That’s an empty square in a parking lot outside “BDA International Business Park,” which includes multiple shops and restaurants within, so that’s obviously not a fab. Problem 2: It appears that Google’s “float over” street grid in Beijing is offset about a quarter mile east from the actual location. Problem 3: Adjusting that marker a quarter-mile west puts it in a cluster of much smaller buildings amidst trees, all of which are obviously too small to be fabs.

Worse still, asking the question differently brings up a completely different address. “Where is ChangXin Memory Technologies located in the Beijing Economic-Technological Development Zone Beijing, China.” Answer:

The specific facility and office address is: No. 51 Jinghai 3rd Road, Courtyard 1, Building 1 (Floors 1-5)Beijing Economic-Technological Development Zone, Beijing, China

Here it can find the road…and nothing else.

Expanding further and further out using aerial views of Beijing in the general vicinity, I find a building complex that might, might qualify, as it has the extensive air-handling equipment that you might find on a fab:

But there’s no guarantee that’s it and not another big factory. (It might also be SMIC’s Beijing fab in the same general area, which Google seems to have trouble finding with any precision as well.)

Likewise, doing a Google image search on “CXMT beijing fab” brings up mostly greenfield artist renderings and pictures of other CXMT buildings, not an actual fab in a Beijing industrial zone. (By the way, the picture of the big building with the curved purple building in the bottom of the pic is an actual fab, but it’s Yangtze Memory Technologies’s NAND flash fab in Wuhan.)

Maybe Google sucks at bringing up information on Chinese locations. In a move I rarely make, let’s ask ChatGPT “What is the exact location of ChangXin Memory Technologies Beijing fab.” Answer:

There is no publicly confirmed street address for the CXMT (ChangXin Memory Technologies) fabrication plant in Beijing….

As of today, CXMT has not publicly disclosed the precise street address or GPS coordinates of the Beijing fab. Unlike many semiconductor companies that publish corporate campus addresses, this facility appears to be intentionally identified only at the industrial park level in public documentation.

As the kids say today, “Super sus, bro.”

What does CXMT’s own website say about their Beijing fab?

“2020.06 CXMT Fab C Project kicked off”
“2022.01 Operation began on Fab C1 (Beijing) trial production line”

OK, now I’m really suspicious. You don’t go from starting work on a fab to pilot production in 18 months, not even in China. Even without waiting for an ASML EUV stepper that they can’t get, the lead time for all the fab equipment alone is likely to be longer that.

To be fair, I couldn’t use Google maps to find their Hefei fab(s) either, but that one does come up in Apple Maps if you search for “Changxin Cunchu Technology Co., Ltd.” And scrolling around Google Maps does bring up some large buildings in that precise location, so I am willing to believe CXMT has one or more fabs there. But searching on “Changxin Cunchu Technology Co. Beijing China” brings up nothing on Apple Maps, and Google automatically changes the name to something else that isn’t a semiconductor company.

Remember, previous Chinese semiconductor companies have proven to be smoke and mirrors and shell games all the way down. Right now, I question whether CXMT even has a fab in Beijing. That doesn’t mean they don’t. The map is not the territory, and Google is obviously having problems with both their mapping and AI technology when it comes to China. (And this is also a lesson, once again, that AI can’t distinguish between truth and falsehood, as “garbage in, garbage out” still applies, and the Internet-based datasets the AI hyperscalers have been using to feed their baby Frankenstein monsters are full of garbage.) But CXMT would not be the first Chinese company to lie about their business to con the gullible. Or even the thousandth.

Anyway, if you have an actual reliable source (not a news article based on a tweet based on press releases) that CXMT has an actual fab in Beijing, feel free to share it in the comments below.

Tags:, , , , , , , , , , , , ,
Posted in Communism, Technology | 5 Comments »

RAM Cartel? I Seriously Doubt It.

Wednesday, July 1st, 2026

A class action lawsuit has been filed against memory giants Samsung, SK Hynix, and Micron alleging the nefarious, cartel like action of…making the products with the highest profit margins.

The world’s biggest memory chip makers are once again facing accusations of manipulating prices.

A class-action lawsuit filed on Thursday, June 25, in a California federal court alleges that Samsung Electronics (SSNLF), SK Hynix (SKHY), and Micron Technology (MU) coordinated to restrict DRAM supply and push prices sharply higher during the AI boom.

The complaint, filed in the US District Court for the Northern District of California under case number 3:26-cv-06345, claims the companies reduced production of traditional DDR3 and DDR4 memory while shifting capacity toward high-margin AI memory products such as HBM chips used in data centers.

Not to mention DDR5.

However, the companies have not been found liable for now, and no trial date is set.

According to the lawsuit, DRAM prices have surged nearly 500%-700% over the past four years, reported Time of India. Plaintiffs argued that in a competitive market, rising prices should attract more supply, but production cuts continued instead.

Snip.

According to Jefferies, memory prices could rise another 40%-50% next quarter and 30%-40% more in the following quarter, reported analysts like Bull Theory on X, with normalization unlikely before 2028. The rising memory costs are already filtering into consumer electronics prices worldwide.

Does this situation suck if you’re trying to buy or build a new PC with lots of RAM? Absolutely. But there’s no nefarious market coordination at work among those big three, just the confluence of a variety of market trends. So let’s break it down:

  • Manufacturers switching production from a less profitable product to a more profitable product isn’t some nefarious conspiracy, it’s how the market works. If they’re getting premium pricing for HBM memory that sells out instantly for the AI bubble, that’s what they’re going to produce. A whole lot of tech companies depended on the spot market for RAM because it gave them more flexibility and costs savings, but now it’s biting them in the ass. Their lack of foresight does not indicate a conspiracy or market failure.
  • Why are there only three big RAM manufacturers? Because a whole lot of other companies dropped out of the market because the game became too expensive to play. RAM makes money hand-over-fist during boom times (like now), but barely breaks even during busts. A whole lot of different companies used to produce memory, Intel and Texas Instruments among them. Remember when Japan Inc. was going to take over the world and the Japanese semiconductor giants (NEC, Toshibu, Fujitsu, Hitachi, etc.) were accused (with some justification) of dumping RAM below cost to capture market share with the backing of state agency MITI? None of those Japanese giants are in RAM any more because, in the wake of the Japanese asset bubble busting in 1991, building new state-of-the-art fabs that doubled in price every four years became a game too expensive for them to play.
  • Rising prices should attract more supply, but it takes about three years and costs about $25 billion to build a state-of-the-art fab. Because standard memory technology still has a capacitance limit, you don’t necessarily need an under-10nm fab, so maybe you can spend a bit less, but you’re still spending over $10 billion on a fab, and you probably still need an ASML EUV stepper, though not the very latest one.
  • And indeed, Samsung, SK Hynix, and Micron all have two new fabs each in the pipeline scheduled to come online this year through 2028. The Micron and SK Hynix fabs will both be dedicated to producing memory. As for Samsung (which has a lot of fingers in a lot of semiconductor pies), I would guess their newest South Korean fab will be dedicated to memory, while their 4-5nm Taylor, Texas fab will not. Building new fabs are not the actions of monopolists who want to artificially constrain supply.
  • Indeed, the “they’re artificially constraining supply” nonsense suggests that they’re producing fewer memory chip than they could otherwise, and that’s just not how the industry works. Fab production lines run 24/7/365 (indeed, they pay technicians triple to work Christmas), because every hour a modern fab is down they’re losing millions in lost profit.
  • Building new fabs is still a risky bet, because the industry is extremely cyclical. No matter how furious the boom now, the next bust is always around the corner. Back when I was working at Applied Materials, the cycle was described as trains linked together with slinkys. First software takes off, then hardware gets yanked along, then the chip manufacturers get yanked, and then, finally, semiconductor equipment manufacturers get yanked into motion, and shortly after that happens, the bust hits the front of the train, and the trailing cars all crash into each other. (The standing joke at Applied Materials was that you could tell the bust was on the very moment the company broke ground on a new manufacturing facility.) Build a new $25 billion fab at the wrong part of the cycle and it could take a company much longer to amortize it than they expected. That’s why so many companies switched to the foundry model.
  • Speaking of foundries, could they be a solution to the memory crunch? Potentially, but there you’re running into the same AI boom-induced wafer start constraints that plague the memory sector. TSMC is fabbing AI chips for Nvidia (and most of its competitors) as fast as it possibly can. Maybe they can profitably book runs on slightly older (but not “mature”) TSMC fabs, but they’re still competing with every other fabless company supporting the AI build-out for the same wafer starts. A whole lot of different silicon goes into a data center.
  • Could an existing semiconductor manufacturer jump into the existing space? Yes, and in fact Intel has announced plans to do just that, though evidently with their own proprietary, next gen “Z-Angle Memory (ZAM),” which isn’t going to do squat to relieve this year’s DDR3/4/5 shortage. Still, they have enough slightly trailing edge fabs to do it, though Intel has had trouble executing at speed in the past.
  • Could another company jump into the semiconductor fab race as an integrated device manufacturer for memory? Risky but possible. Someone like Apple could decide that memory shortages are an existential threat to its business model and spend the tens of billions to get into the game. And indeed, Apple is already spending some $500 billion to reshore its supply chain back into the US, so that would fit right in. Apple could potentially contract with TSMC (or even Micron) to build and run a memory fab. (Samsung is a trickier proposition, since the two are fierce competitors as the biggest smartphone manufacturers in the world, but there’s still a lot of “cooperatition” between the two, so it’s not beyond the realm of possibility.) But the three year lead time still applies.
  • Entire tech boom and bust cycles have come and gone in an era in which RAM is cheap and plentiful, a situation people have come to think of as “normal.” Just as with higher credit rates, a whole lot of business models that were viable in an era of cheap memory are suddenly going to stop being so in an era of scarcity. Some companies will be able to raise prices and remain profitable, and others won’t. Not everyone will be hit, as a lot of embedded devices use older types of memory that hasn’t gone through the roof. There are all sorts of older fabs churning out older types of memory that aren’t relevant to this discussion.
  • The idea that Samsung and SK Hynix are colluding is particularly laughable, as the two Korean chaebol backing SK Hynix (Hyundai and LG (AKA Lucky Goldstar)) both hate rival Samsung with a passion.

    • The current shortage, as painful as it is to so many, isn’t the result of a nefarious cartel, it’s just the free market working like it always does at the interface between supply and demand. It’s just that cutting-edge semiconductor supply has a whole lot more lead-time constraints that most other economic sectors.

    The AI Bubble seems considerably worse than the Dotcom Bubble (which was only partially about the Internet; updating hardware and software to avoid the Y2K bug also drove a lot of spending in the same timeframe), and its inevitable bursting (or just deflating) is going to relieve pressure on everyone else that needs 10nm or smaller wafer starts.

    But there’s no telling exactly when that will be.

    Tags:, , , , , , , , , , , , , , ,
    Posted in Economics, Technology, Texas | 8 Comments »

    AI News Roundup For February 5, 2026

    Thursday, February 5th, 2026

    A bunch of AI-related news has popped up this week, so let’s do a roundup.

  • Some AI companies are complaining that TSMC is killing the AI boom by not expanding rapidly enough:

    Asianometry notes that TSMC’s caution at expanding is amply justified by the boom-and-bust nature of the semiconductor industry:

    • “I’m hearing many similar views in the Silicon Valley Borg that TSMC is the break or limiter on the AI boom, as if they’re the reason why we don’t have AGI yet. Because they didn’t and still don’t believe.”

    • “If we can ever say that a company that spent $41 billion on capital expenditure in 2025, with another $53 to $56 billion in 2026 planned, is sitting on its hands, doing nothing.”
    • “TSMC having 90% share of the AI chip market looks pretty unhealthy. That should go down and it will. Samsung seems to be doing well so far.”
    • “The cold, hard reality is that shortages are a fact of life in semiconductors, as are horrific gluts.”
    • “What we are flippantly labeling as TSMC we really mean is the AI supply chain. And that supply chain is as complicated as you can possibly imagine. Like an iceberg, it looks big enough on the surface of the water, but goes way far deeper underneath. TSMC has thousands of suppliers in two categories: Equipment like the famed ASML lithography tools and materials like photoresist, silicon wafers, acid etch gases and so on. These are not generalized tools and materials. They are not fungeible like AWS compute units.”
    • “And then there are the memory guys. You cannot ship an AI system without memory. DRAM and NAND. Nvidia’s AI chips use a special form of DRAM called high bandwidth memory, and they use quite a lot of it. The memory industry is just as consolidated as the logic industry, with the major players being Samsung, SK Hynix and Micron.”
    • “The chip guys are last to know when the party is getting started, but first they get batoned in the face when the police shut things down.”
    • He points out that semiconductor manufacturers have log supply chains. He uses a different metaphor (the beer distribution game, or a bullwhip), but back when I was working at Applied Materials, it was described as trains linked together with slinkys. First software takes off, then hardware gets yanked along, then the chip manufacturers get yanked, and then, finally, semiconductor equipment manufacturers get yanked into motion, and shortly after that happens, the bust hits the front of the train, and the trailing cars all crash into each other. It’s a regular boom/bust cycle.
    • “From 1961 to 2006, electronics consumption in the United States grew positively but with wild volatility swings between 0 to 20%. But for the semiconductor makers, that translates to swings anywhere from 20% to 40%. And for the equipment makers, it is amplified even more, plus or minus 60%. The whip hits particularly hard in the semiconductor industry because of the industry’s long lead times. It takes 4.5 months to fabricate and package a chip. It takes 18 months to 2 years to build a fab. Meaning from shovels down to producing chips, and it takes 12 to 18 months to produce and install something like an EUV machine into the fab. Another 6 months before that machine actually starts patterning wafers.”
    • “Long lead times mean having to make very long demand forecasts, which leads to extreme volatility swings during up and downturns even if those up or downturns are relatively small.” People forget that in 1998, during the time we now think of as the DotCom Boom, there was a small semiconductor downturn that had Applied Materials forcing employees to take unpaid leave.
    • “ASML just reported 2025 earnings, and we see the bullwhip in full effect. TSMC raised capital expenditure 35% but ASML announced €13.2 billion of net new bookings. Analysts had expected just €6.32 billion. This is because ASML collected orders not just from TSMC, but also Samsung, Intel and the memory guys. When it rains it pours, right? Again, this is why I fear that another AI foundry would not mean our compute shortage is solved, because ultimately, when those foundries start scaling their capacity, they all go to the same suppliers.”
    • He goes over how car manufacturers cancelled orders during Flu Manchu, and then scrambled when the economy took off afterwards. “TSMC was trying to discern between double booked orders and real demand, which is not an uncommon experience for them. Customers lie about their own demand all the time, or at least we can say that they are eternally optimistic. TSMC tried to respond in 2022. The Taiwanese giant poured $36 billion into capital expenditure. They went to their suppliers and pushed like no tomorrow.”
    • “It turned out those customers really were double booking orders and artificially inflating demand. When the macro environment turned in 2022, the automotive, smartphone, and PC chips that were so hot during the COVID era fell out of vogue and customers started cutting orders.”
    • “Meanwhile, deeper down in the supply chain, TSMC and the rest of the semiconductor industry were getting bullwhipped by COVID hangover. Utilization at TSMC’s multi-billion dollar N7 fabs crashed, Semi analysis wrote in April 2023. Now, Semi analysis data indicates that the 7nm utilization rates were below 70% in Q1. Furthermore, Q2 gets even worse with 7nm utilization rates falling to below 60%. This is primarily due to weakness in both smartphones and PCs, but there is a broader weakness in most segments. A fab’s break even utilization rates are about 60% to 70%. So those N7 Taichung fabs were taking financial losses potentially on the order of hundreds of millions, maybe even billions. The financial burdens of low utilization are another reason why I’m skeptical another AI foundry could have rushed into the AI chip fray to save the day.”
    • He says that Intel incurred losses during this period due to an unnecessary fab expansion, which is probably true, but that was a secondary factor next to their longer running problem of getting their process wrong.
    • “ChatGPT was released in November 2022, and that kicked off a massive increase in capex amongst the hyperscalers in particular, but it sure seems like TSMC didn’t buy the hype. That lack of increased investment earlier this decade is why there is a shortage today and is why TSMC has been a de facto break on the AI buildout/bubble.”
    • “I recall news in mid 2024 of TSMC struggling with CoWoS capacity bottlenecks and yield problems, including one design issue that caused cracks in the Nvidia chips packaging.” CoWoS is Chip on Wafer on Substrate, which involves fabbing an interposer as a substrate for faster connections between your processing chips and memory.
    • “I also recall news in late 2024 noting how the vendors in charge of making the server racks for Nvidia’s Blackwell servers struggled with overheating, liquid cooling leaks, software bugs, and connectivity issues. Such technical difficulties delayed server deployment until early to mid 2025, creating a weird situation for several months where TSMC was pumping out chips that just went into storage. So that gated things, because you don’t scale until you first fix the technical problems.”
    • Then there’s the power-scaling issue, which is a whole ‘nuther can of worms.

  • There’s a lot of talk about a SaaSpocalypse going on thanks to a new AI tool. (SaaS is “Software as a Service.” Instead of hosting your own payroll or sales-tracking or whatever servers, you hire a company that already has cloud software setup to do it and you just tie into that, which can considerably reduce startup costs. A whole lot of successful new tech companies over the last decade plus have been SaaS companies.)

    The software sector was jolted overnight with what analysts are calling a “SaaSpocalypse” — a sudden and severe selloff triggered by new artificial intelligence tools unveiled by US AI startup Anthropic. The episode has sharpened investor fears that AI is no longer merely helping software companies but may now begin replacing them.

    Anthropic has expanded its enterprise AI platform, Claude Cowork, by launching 11 new plugins aimed at automating a wide range of professional tasks. Claude Cowork is an agentic, no-code AI assistant built for corporate users, allowing companies to automate workflows without writing software. The new plugins are designed to handle tasks across legal, sales, marketing and data analysis functions. The most recent addition is Anthropic’s Claude Legal agent, which can perform routine legal work such as document and contract review, and compliance checks.

    Anthropic has said that the tool does not provide legal advice and that all AI-generated outputs must be reviewed by licensed attorneys. Even so, the breadth of automation signals a step change in how much white-collar work AI systems can now perform.

    Here are the current plugins for Claude Cowork:

    • Productivity — Manage tasks, calendars, daily workflows, and personal context
    • Enterprise search — Find information across your company’s tools and docs
    • Plugin Create/Customize — Create and customize new plugins from scratch
    • Sales — Research prospects, prep deals, and follow your sales process
    • Finance — Analyze financials, build models, and track key metrics
    • Data — Query, visualize, and interpret datasets
    • Legal — Review documents, flag risks, and track compliance
    • Marketing — Draft content, plan campaigns, and manage launches
    • Customer support — Triage issues, draft responses, and surface solutions
    • Product management — Write specs, prioritize roadmaps, and track progress
    • Biology research — Search literature, analyze results, and plan experiments

    A lot of those are already automated elsewhere, but I suspect a lot accountants and paralegals just felt a goose strut across their grave. On the other hand, who is really going to turn over, say, Accounts Payable to an AI? One glitch, and your entire bank account is drained…

    If it works (a big if, give so many AIs are prone to hallucinations), this is potentially good news for Anthropic and the companies using their tools, and bad for SaaS companies and the employees currently doing those jobs.

    I note there’s no plugin for technical writing…yet.

  • Google/Alphabet just reported $400 billion in earnings in 2025. CEO Sundar Pichai:

    And Google Cloud ended 2025 at an annual run rate of over $70 billion, representing a wide breadth of customers, driven by demand for AI products.

    We’re seeing our AI investments and infrastructure drive revenue and growth across the board. To meet customer demand and capitalize on the growing opportunities we have ahead of us, our 2026 CapEx investments are anticipated to be in the range of $175 to $185 billion.”

  • Remember how Nvidia was going to invest $100 billion in OpenAI? Yeah, not so much.

    In September 2025, Nvidia and OpenAI announced a letter of intent for Nvidia to invest up to $100 billion in OpenAI’s AI infrastructure. At the time, the companies said they expected to finalize details “in the coming weeks.” Five months later, no deal has closed, Nvidia’s CEO now says the $100 billion figure was “never a commitment,” and Reuters reports that OpenAI has been quietly seeking alternatives to Nvidia chips since last year.

    Reuters also wrote that OpenAI is unsatisfied with the speed of some Nvidia chips for inference tasks, citing eight sources familiar with the matter. Inference is the process by which a trained AI model generates responses to user queries. According to the report, the issue became apparent in OpenAI’s Codex, an AI code-generation tool. OpenAI staff reportedly attributed some of Codex’s performance limitations to Nvidia’s GPU-based hardware.

    After the Reuters story published and Nvidia’s stock price took a dive, Nvidia and OpenAI have tried to smooth things over publicly. OpenAI CEO Sam Altman posted on X: “We love working with NVIDIA and they make the best AI chips in the world. We hope to be a gigantic customer for a very long time. I don’t get where all this insanity is coming from.”

  • You know who’s not winning the AI war? Microsoft.

    Microsoft’s Copilot chatbot has become central to its artificial-intelligence strategy as the company’s close partnership with OpenAI diminishes. But the effort to build it up as a ChatGPT alternative has been tough going.

    Remember, Copilot is the AI that wants to take pictures of your desktop every few seconds. Golly, can’t imagine why it’s unpopular..

    Confusing brand positioning and interoperability problems have frustrated users, current and former employees who have worked on Microsoft’s AI products said.

    Interoperability problems? With a Microsoft product?

    Only a small proportion of subscribers to Microsoft’s enterprise suite use Copilot, and the percentage who favor it over Google’s Gemini or other tools has decreased in recent months, according to data reviewed by the Journal.

    The stakes are high for Microsoft because Copilot is core to a push by Chief Executive Satya Nadella to transform Microsoft into an AI-first company, much as he transformed it into a cloud-first company around a decade ago. Copilot is one of Nadella’s top priorities, current and former executives said.

    Microsoft shares tumbled after its earnings report last week sparked investor concern that growth in its most important unit, the Azure cloud-computing business, is slowing, and that its AI business is reliant on OpenAI while Copilot remains unproven. Shares fell nearly 3% Tuesday amid a slide in software stocks prompted by fresh concerns that AI tools will make enterprise subscriptions less necessary.

    For other AI companies, we merely suspect they’re evil. For Microsoft (and Google), we already know they’re evil…

    • Tags:, , , , , , , , , , , , , , , , , , , , , , , , , , , ,
    Posted in Economics, Technology, video | 4 Comments »

    Instant Analysis: Trump Tariff Effects On Semiconductors

    Wednesday, April 2nd, 2025

    President Trump announced his tariffs on countries, especially those that tariff goods from the United States.

    President Donald Trump on Wednesday imposed sweeping new tariffs on all imported goods and unveiled a detailed list of reciprocal duties targeting more than 60 countries, asserting that the move is necessary to combat trade imbalances and restore U.S. manufacturing.

    “This is Liberation Day,” Trump said during a Rose Garden ceremony, holding up a printed chart of countries and their new tariff rates. “For decades, our country has been looted, pillaged, raped and plundered by nations near and far, both friend and foe alike.”

    The tariffs, which he described as “reciprocal,” fulfill a key campaign pledge and are aimed at pressuring trade partners to lower their own barriers. The administration expects the new rates to remain in place until the U.S. narrows a $1.2 trillion trade imbalance recorded last year.

    But the extensive list of tariffs also threatens to upend the U.S. economy, as many — but not all — economists say they amount to taxes on American companies that will be passed down to consumers.

    Trump held up a chart while speaking at the White House, showing the United States would charge a 34 percent tax on imports from China, a 20 percent tax on imports from the European Union, 25 percent on South Korea, 24 percent on Japan and 32 percent on Taiwan.

    The centerpiece of the announcement is a 10 percent universal baseline tariff on all imports, effective immediately. For instance, Chinese imports are now subject to cascading tariffs of 10, 20 and 34 percent, for a total of 54 percent.

    In addition, Trump’s administration imposed country-specific reciprocal tariffs on nations it accuses of unfair trade practices — including India, Vietnam, and the European Union, in adding to China. The rates are calibrated at approximately half the rate those countries impose on U.S. goods.

    For example, China, which Trump said charges 67 percent in tariffs on U.S. goods when factoring in non-tariff barriers, will now face a 34 percent reciprocal tariff under the new system, in addition to the 10 percent baseline tariff and the 20 percent tariffs already in effect. Vietnam, assessed at 90 percent, will face a 46 percent tariff; India at 52 percent will now see 26 percent duties; and the EU, which imposes 39 percent, will be met with a 20 percent response, according to the White House chart.

    This is a “devil in the details” issue that has a lot of ramifications depending on how the directives are written. But several of those countries are big players in semiconductors, so here’s a quick and dirty look at winners and losers if those tariffs stay in place a significant amount of time.

    The main countries here, along with the reciprocal tariffs being applied to them:

  • Taiwan (32%)
  • South Korea (25%)
  • China (34%)
  • European Union (not a country, but they play one on TV) (20%)
  • Japan (24%)
  • Singapore (10%)
  • Israel (17%)

    • Save a few smaller, older fabs here and there, that’s pretty much 99% of semiconductor manufacturing, though Vietnam (46%) and the Philippines (17%) do a lot of semiconductor package assembly work, and the tariffs may apply to them, depending on wording.

    So let’s look at the business Losers and Winners in the space. (Note: You might find this post useful, as it defines some of the semiconductor industry terms used here.)

    Losers

  • TSMC: As the world’s biggest and most important chip foundry, the Taiwanese tariffs will hit TSMC hard. Their U.S. fab in Arizona isn’t ready for production yet, so all their chips will (theoretically) get hit with tariffs, assuming Trump doesn’t grant them a waiver because they’re already constructing a plant. But if they do go into effect, possibly even more heavily impacted will be:
  • TSMC customers, including Apple, Nvidia and AMD. All three get their very highest-end, cutting edge, sub-10nm chips fabbed there. For Apple, the M-series and A-series chips made there form the heart of all their Macs and iPhones. Likewise, Nvidia gets its highest end GPU/AI/etc. chips fabbed by TSMC. AMD’s most powerful CPU’s are also fabbed by TSMC, though some lower end chips are made elsewhere (like GlobalFoundries).
  • Tokyo Electron: Japan’s biggest semiconductor equipment manufacturer assembles pretty much all their equipment in their home country. 24% tariffs may make their equipment uneconomical compared to rivals Applied Materials and LAM Research.
  • South Korean DRAM manufacturers Samsung and SK Hynix: 25% tariffs will definitely impact sales in a market segment whose overall margins (robust in booms, and barely breaking even during busts) are thinner than others.
  • Every American electronics company that uses DRAM. Which is pretty much every American electronics company.
  • Every American AI boom company. Their data center costs are going up, while those of their foreign competitors are not.
  • Korean flat panel display manufacturers Samsung and LG Semicon, who between them control over 50% of the market.
  • Every American TV and monitor manufacturer, the vast majority of which have their devices manufactured overseas.
  • UMC: They’d fallen woefully behind TSMC for foundry work, and they won’t be winning much additional American business now.
  • Every company trying to build a sub-10nm fab in the U.S., as steppers from Netherlands-based ASML just got more expensive and the competition to obtain them might have increased.
  • Pretty much every fab in China just got more screwed…but they were pretty screwed (and trailing badly) before.
  • American fabless chip startups: Their costs for getting chips to market probably increased.

    • Winners

  • Applied Materials, LAM Research and KLA Tencor. Buying competing Tokyo Electron equipment just got more expensive, and a bunch of companies now have incentives to build fabs in America.
  • Intel: Assuming they’ve finally got their process technology sorted out (a big if), they’re well-positioned to take CPU market share from AMD and to grow their under-performing foundry business.
  • Micron (sort of): As the only American DRAM manufacturer, they can probably earn more per each chip produced domestically. But Micron has a lot of overseas fabs these days, and building new domestic DRAM fabs will take years.
  • GlobalFoundries: The costs of their global competitors just increased, so they can probably win more business for their domestic foundries…if they have the available wafer starts. But they have a lot of foreign fabs as well.
  • Samsung‘s US foundry business. Presumably the wafer starts for their Austin and Taylor fabs will see increased demand.
  • Maybe Texas Instruments, but I’m not sure how much mixed-signal and analog competition they have, and that’s their bread and butter.

    • Neutral

  • ASML: Being in the Netherlands and having TSMC as their biggest customer, you figure they’d be hurt, but no. You can’t get EUV steppers from anyone else, and I get the impression they’re building EUV steppers as fast as they possibly can already. Anyone building a cutting-edge fab will just have to pay more to get them.
  • Tower Semiconductor: Half their foundries are in Israel and half in the U.S., so I figure it’s a wash.

    • That’s my quick and dirty analysis. Of course, Trump is using tariffs like a battering ram to smash foreign tariffs, and if he’s immediately successful, there probably will only be minor hiccups in the global supply chain. But if not, a whole lot of disruption might lie ahead, and it usually takes a minimum of 3-5 years to bring a new fab online.

    Tags:, , , , , , , , , , , , , , , , , , , , , , , , , , , , , ,
    Posted in Austin, Economics, Foreign Policy, Texas | 6 Comments »

    China’s Semiconductor Industry: Shell Games All The Way Down

    Wednesday, April 7th, 2021

    I’ve written about China’s semi-illusory semiconductor businesses before: “In China the question is always how much of that investment is real, and how much is illusion. A lot of those ‘under construction’ fabs never materialize, either unable to attract investors or having their funds magically siphoned off to some other enterprise.” While researching yesterday’s piece on the current semiconductor shortage, I came across this Emily Feng NPR piece on more multi-million dollar shenanigans in that space:

    In 2019, the U.S. sanctioned two major Chinese telecom firms, temporarily cutting them off from a vital supply of semiconductor chips — bits of silicon wafer and microscopic circuitry that help run nearly all our electronic devices.

    Wuhan Hongxin Semiconductor Manufacturing Co. promised a way out, toward self-reliance in the face of increasingly tough U.S. curbs on this technology. The private company once boasted on its website that it would raise a total of $20 billion to churn out 60,000 leading-edge chips a year.

    None of that would come to pass.

    Hongxin’s unfinished plant in the port city of Wuhan now stands abandoned. Its founders have vanished, despite owing contractors and investors billions of yuan.

    The company is one of six multibillion-dollar chip projects to fail in the last two years. Their rise and fall is a cautionary tale in an industry that is flush with state cash but still scarce on expertise — and a preview of the expensive and winding road China will have to take toward semiconductor self-sufficiency, now a national security priority.

    Hongxin Semiconductor began in November 2017 as a joint venture between Wuhan’s Dongxihu district government and a company called Beijing Guangliang Lantu Technology.

    The venture got off to a good start — on paper — but a closer look shows there were a number of issues. One of the co-founders of Guangliang had only finished elementary school and was allegedly using false credentials and a different identity, Cao Shan, according to 36Kr, a Chinese tech news outlet. Another co-founder, Li Xueyen, dabbled in selling Chinese traditional medicine, alcohol and tobacco before starting Hongxin, according to corporate records reviewed by NPR.

    These are not the profiles you look for in semiconductor startup founders.

    The two could not be reached for comment.

    Yeah, I bet.

    To balance out their lack of technical know-how, the Hongxin founders lured in one of Taiwan’s most famous semiconductor engineers, Chiang Shangyi, to serve as director. He left the company in 2020 to become the deputy chairman of China’s Semiconductor Manufacturing International Corp., telling Hong Kong paper South China Morning Post that his time at Hongxin was “a nightmare.” Chiang did not respond to NPR requests for comment.

    Hongxin made headlines in December 2019 when it managed to buy an older model lithography machine made by Dutch company ASML, despite American lobbying to prevent its sale to the Chinese chipmakers.

    OK, on the face of it that sounds pretty impressive. If you want to have a cutting edge fab, you have to have one of ASML’s top of the line Extreme Ultraviolet (EUV) steppers. In almost every other segment in the semiconductor equipment market, there’s competition between the three big players (Applied Materials, LAM Research and Tokyo Electron) and occasionally other companies (like Axcelis for ion implanters). But while you might be able to get away with lesser Deep Ultraviolet (DUV) lithography machines from Nikon or Canon for some tasks, for the smallest features on cutting edge 7 and 5nm nodes, you simply can’t do without an ASML EUV stepper. (More background here.)

    Well, guess what? The vaunted ASML tool Hongxin bought is apparently an older 1980 model (presumably this one, which dates from 2015, not 1980) which is DUV, not EUV.

    Back to the NPR piece.

    ASML sold the multimillion dollar piece of equipment — used to etch semiconductors — because of Jiang’s top-notch reputation, according to two people familiar with the sale who were not authorized to speak publicly about it. ASML declined to comment.

    Feng (or her editors) goofed here. ASML makes lithography machines, not etch tools.

    Hongxin’s timing was opportune. Chinese chip companies still rely heavily on European, American and Japanese technology — much of which, in turn, relies on American intellectual property, which the U.S. appears determined to keep out of Chinese hands. China’s semiconductor demand continues to surge beyond what it can supply itself; trade data show that in 2019, Beijing imported around $350 billion worth in chips.

    Given that reliance, China’s central and local governments have been pumping money into the sector to accelerate domestic chip design and manufacturing. The country’s latest five-year economic planning document released in March identifies integrated circuits — semiconductors — as a priority sector for research and development funding.

    When governments starts pumping big money into private companies, you can be sure multiple scams are never far behind.

    The all-out approach has notched achievements. Successful chip design companies such as Cambricon and Huawei’s HiSilicon have allowed Huawei to replace some of its U.S.-designed chips in its mobile phones.

    Cambricon and HiSilicon are both fabless design houses, and both get their chips fabbed at foundries like TSMC. Huawei is one of the largest electronics companies in the world, with over $100 billion in annual sales, and they don’t own their own fab.

    Not far from Hongxin is Yangtze Memory Technologies Co. (YMTC), a partially state-owned company that plans to double its output of memory chips to overtake South Korea’s Samsung and SK Hynix, which currently dominate production.

    Memory is a tough business. SK Hynix exists because Hyundai and LG (aka Lucky Goldstar), two huge Korean chaebols who hate each other only slightly less than rival Samsung, found the sledding too tough to go alone and had to combine their respective semiconductor operations to survive. Memory makes money hand-over-fist in boom times, but barely breaks even during busts. It’s less technically demanding than some other semiconductor segments, so China could conceivably make some headway there.

    YMTC is a subsidiary of Tsinghua Unigroup, a wholly owned business unit of Tsinghua University. Hu Haifeng, Communist Party secretary of Tsinghua Holdings, is the son of Hu Jintao, former CCP General Secretary and President of the People’s Republic of China.

    Hongxin sought to capitalize on this momentum. It rented a discreet office on the 25th floor of Wuhan’s Dongxihu district government headquarters.

    “Cao” and his partners promised to pitch in 1.8 billion yuan ($276 million) in investment on top of 200 million yuan ($30.7 million) in starting funds from Dongxihu district.

    Wuhan’s city government was, around the same time, also beginning construction on a cybersecurity park to provide office and residential space for technology businesses, and it was looking for a flagship company to anchor the complex. In 2018 and 2019, the city named Hongxin its most important “critical construction project” and the company began building its factory next door.

    As early as late 2019, even while Hongxin was being lauded by Chinese media for securing an ASML machine, several Wuhan-based construction crews were scrambling to get paid for millions of dollars of work for Hongxin.

    “Four months ago, [Hongxin’s] payments to us started to be short, and now we are missing 18 million yuan [$2.76 million],” one contractor, Lu Haitao told another, Wang Liyun in December 2019, according to phone recordings NPR obtained. Wang confirmed the authenticity of the recordings when reached by phone. Lu did not respond to several texts and calls from NPR. Wuhan’s municipal government did not respond to a request for comment.

    Meanwhile, two other semiconductor companies — Tacoma Semiconductor Technology Co. Ltd. and Dehuai Semiconductor Technology Co. Ltd. — were also running out of cash.

    Tacoma was over 350 miles from Hongxin along the Yangtze river, in the port city of Nanjing. There, the Taiwanese entrepreneur Joseph Lee had initially found a welcome harbor for his own ambitions, starting Tacoma in the city in 2015. He pledged to raise $3 billion to make wafer chips, with consultation from Israeli company Tower Semiconductor (formerly TowerJazz). Tower declined to comment for this story.

    Lee continued pitching other local governments. In 2016, he co-founded a second company in Jiangsu province’s Huai’an city, named Dehuai Semiconductor. (Lee sold his stake the same year, citing a clash in vision with the firm’s other managers.)

    In 2017, Lee invited Chinese media to tour Tacoma’s facilities, declaring the company had somehow scored 200 million yuan ($30.7 million) in sales. Tacoma had yet to even finish construction on its manufacturing facilities.

    Lee initially agreed to an NPR interview for this story but later retracted it, citing state pressure. “Officials have told me not to talk to the media,” he said by text.

    Yeah, I bet.

    By 2018, Tacoma’s employees were blasting an online forum run by the Nanjing mayor’s office with complaints about unpaid salaries. Chinese corporate records show at least 50 legal complaints have been filed against Tacoma in provincial court, all seeking to recoup construction costs or unpaid wages. Lee disputes owing employees 20 million yuan in unpaid wages.

    “Real or fake, the truth is in the hearts of the people,” Lee wrote shortly after these allegations, on Wechat, the Chinese messaging app, and cited a verse from the New Testament: “Now faith is the certainty of things hoped for, a proof of things not seen.”

    Citing bible verse when rumbled for his scam. Classic.

    Hongxin, Tacoma and Dehuai were able to secure billions of yuan in state funding on the condition they would match that with investment of their own — a commitment that never materialized. Tacoma eventually raised only a fraction — 250 million out of 2.5 billion yuan — of what it promised.

    “We never imagined that when our cash flow dried up, we would not be able to find new [cash flow sources], that we would get in so deep,” he told Japanese broadcaster NHK this March.

    And this is the problem with doing business in China in general: it’s shell games all the way down. At lot of times, loans and investments are siphoned through four or five different entities from the purposes for which they were originally obtained. Everyone’s trying to get rich, and they hope to survive on smoke and mirrors long enough to get profitable. Imagine if Kleiner Perkins invested $25 million in a software startup, only to find that money was spent on a noodle shop, a used car dealership and a golf club manufacturer.

    Sometimes it works. You can build a company on margin, get profitable quickly, and be paying off investors and contractors before anyone realizes how shaky the entire enterprise is.

    But you can’t do that with semiconductor manufacturing. The startup costs are simply too high, easily in the billions. Very, very few companies can afford to be in a game that expensive. China’s two biggest semiconductor manufacturing success stories, SMIC and Tsinghua Unigroup, all have have CCP direct government investment.

    In this game, little hucksters working the margins have no chance.

    Tags:, , , , , , , , , , , , , , , , , , , , , , , , , ,
    Posted in Communism, Foreign Policy | 14 Comments »