Huawei tells AI chip buyers to source their own memory as HBM smuggling prices hit $4,000

The arrangement concerns High-Bandwidth Memory, or HBM. This is the stacked memory that sits beside an AI accelerator and feeds it data. Huawei's Chinese suppliers cannot produce enough of it. U.S. export rules also block direct purchases from Samsung and SK Hynix, which together control more than 80 percent of the global market, according to The Decoder's 3 November report.

So the customer is told to find the memory elsewhere.

What "elsewhere" means is a smuggling chain built around a loophole in U.S. export rules. Smugglers bundle HBM with cheap server processors whose performance falls below the export threshold. They ship the packages to China through Hong Kong, then separate the memory from the processors in workshops there. The Decoder describes the process as fragile, with a high error rate, but profitable enough to keep running. The report cites The Information for the core claim that Huawei is pushing procurement onto customers, and adds pricing detail from a broker in Shenzhen.

What the memory costs

That broker puts smuggled HBM3E at around $4,000 per unit, up from a few hundred dollars a year earlier. Each AI chip needs six to eight units, so the memory alone can reach $32,000 per chip. Huawei plans to launch the Ascend 960DT early next year. It says the chip will double the performance of its current chips, according to the same report.

The smuggling routes are not the only supply. Older Nvidia GPUs stripped for parts feed the market. So does excess inventory from chip designers who scrapped projects or cut orders.

Governments are counting the leakage. South Korea's customs authority had uncovered trade violations worth $560 million by May, already above the previous full-year record, The Decoder reported. A third of that total involved strategic goods such as chips exported without proper authorization or with false declarations.

The Huawei story is the sharpest illustration this week of a wider squeeze. The hardware that AI depends on is scarce, expensive and increasingly moved around the rules rather than through them. The same shortage shapes the packaging research below. It also shapes the labour market debate that has dominated coverage over the past few days.

Two ways to cut the footprint

On 2 November, SemiEngineering flagged a paper from imec and Vrije Universiteit Brussel. The paper models the embodied emissions of scaling a chiplet-based GPU. The researchers doubled the number of compute units and the HBM capacity, then compared 2.5D and 3D packaging. 3D scaling produced an 89 percent increase in embodied emissions, against 141 percent for 2.5D. The authors present this as an argument for 3D integration in low-carbon AI. The paper is an arXiv preprint, arXiv:2610.32830, so it has not been through peer review.

Memory scarcity also shows up in the electric vehicle battery data. Global EV battery usage reached 994.5 GWh in January to September, up 19.7 percent year on year, according to SNE Research figures published on Monday and reported by CnEVPost. CATL held 39.1 percent and BYD 15.2 percent, a combined 54.3 percent. September set a monthly record at 145.1 GWh, though year-on-year growth slowed to 15.8 percent, below the cumulative rate, CnEVPost reported on 2 November.

Those numbers are about cars, not data centres. But they describe the same industrial base that AI hardware competes with for components and capital.

The office that is not an office

The labour side of the AI question looks different depending on which measurement you pick up. On 2 November, CNBC reported on a New York Fed survey of 1,835 college-educated workers, conducted in September. It found that 48.8 percent had no in-person meetings on their most recent office day. The paper, by NY Fed research economist Natalia Emanuel and University of Virginia economics professor Emma Harrington, concludes that "a staggering share of the workforce experiences this form of in-office remote work."

"Amazingly, 37% of respondents reported that they had at least one meeting that was virtual, even though all the participants were in the same office that day," the researchers write, per CNBC.

Roughly one in four workers is now fully remote, according to U.S. Bureau of Labor Statistics data cited by the NY Fed. The researchers say this figure likely understates how remote from colleagues people feel. That finding sits awkwardly beside the return-to-office arguments companies have made about collaboration. It is also a reminder that the tools credited with productivity gains reorganise who talks to whom.

On the AI and jobs question specifically, the evidence does not settle the argument. The recent headlines gathered for context point in several directions at once. Goldman Sachs research says displaced workers take longer to find new roles. A Statistics Canada study finds little evidence that AI explains a decline in youth hiring. China has an 18-point plan to monitor job displacement risk. None of those items is in the source set used here, so treat them as pointers, not findings.

What the dated material does support is narrower. The New York Fed survey is a measurement of isolation, not of automation. The SNE Research data is a measurement of battery installations. The imec paper is a measurement of embodied emissions. Each is a single instrument reading. The temptation to stack them into one verdict on AI and work is exactly the temptation to resist.

Meanwhile the memory trade keeps moving. Huawei's customers have been handed a procurement problem that no vendor can solve for them. The price of solving it is set by brokers in Shenzhen and customs officers in Seoul.