At it's core, running a top-tier LLM means linear load of trillion+ weights from RAM for every token. The cost of this operation absolutely dwarfs the cost of doing the actual math. If you are only serving a single user, you can only amortize this over a handful of token streams, times like 2-4 for speculative decoding. If you are running in a datacenter, you can batch thousands of streams, and get correspondingly much more tokens out of the same expenditure of energy and hardware.

The current-gen hardware is expected to have only 2-3 years of useful life. New tech, and especially new architecture is going to make inference much faster and more efficient in the near future, and the big vendors are going to be more limited by datacenter floor space and power/cooling, so that the hardware will be replaced very quickly by better models.

also, power and cooling agreements are usually 10y too.

Expected by whom and why?

Google hasn't retired any of its now ten-year-old TPUs, or any of the subsequent generations. Maybe TPUs are special or Google doesn't understand compute economics but either argument requires data.

If your argument is depreciation, you're confusing a tax concept (depreciation) with useful life. Yes, they're supposedly related, but in practice, the relationship is rather loose. (Residential housing depreciates in 28 years yet typically lasts 2-3x as long and 4x is not rare.) Google's V1 TPUs

The actual usable bandwidth was always much lower, somewhere between 80kB/s to a bit over 380kB/s, depending on how you were using the bus, because the cpu needed to fetch its instructions on the same bus. In a way, the bus was usually always running at 1.19MB/s though, because all the basic devices (ROM, RAM, common simple ISA cards) were able to always respond in the basic 4-cycle transfer, it's just that most cycles were spent fetching instructions instead of actual data.

Later machines made the bus faster by increasing width, increasing clocks, and dropping one mandatory wait state at AT for 5.55MB/s theoretical, and dropping the last wait state in the fast clones for 8.33MB/s theoretical.

This excludes the R&D costs, including training, where most of their costs lie.

The basic truth of LLM economics is that inference is cheap. The NRE cost to get where you can get premium prices for inference is sky-high. This is part of why the AI companies want to drive the scale up so much.

The API Gross margin is even higher.

The purpose of the high API prices is, in addition to trying to get more money, to anchor the high price, and then make the subscriptions seem like a great deal. But inference is just not that expensive, not when provided at scale and with huge batches.

You can see this with the various Chinese providers doing inference on open models, which are unlikely to cost less to run than the models at the frontier AI labs. A single 8-GPU GB300 cluster doing high-throughput batch inference on Kimi K3.1 Pro, which is the largest open weights model available, can do something like 300 000 tokens per second, that is, 10T tokens per year. That cluster, with current insane inflated prices, costs something like $1M to buy, and $50k to run per year in electricity. At $10/Mtok, that pays itself back in 4 days. Do you think that that's a reasonable price?

The subscriptions were probably priced with reasonable positive margins.

The huge expenses are in training and pushing the models forwards, with the gigantic RLAIF runs we know they are doing. Then once you got the model, you can scale up inference at low cost and want to serve as many people as possible.

The uppermost parts of the atmosphere are very dynamic, with changes in solar activity causing them to contract and expand on a scale of hours. And this is very significant for the amount of drag felt by re-entering objects. Because of this, if you slowly deorbit something naturally through drag, you are not going to be able to pick the destination, or even know where it ends up until very close to it falling down. So instead you intentionally lower the perigee so that it deorbits in a controlled manner.

Thanks for the correction!

Krzanich "optimized" the company for profit, by cutting out things he considered to be excess. Including a lot of their irreplaceable validation & verification talent, that AMD instantly snapped up. The lack of validation resources is directly responsible for the failures of many key Intel products during his tenure and later, because major bugs that could have been spotted earlier slipped through, and then resulted in horrible delays and even cancellations. He then spent the surplus of the company on 30 major acquisitions, of which only two were in any reasonable degree successful.

There are CEOs who have destroyed more value than Krzanich, but not many.

Presumably they will publish it with sufficient fanfare once patches are applied widely enough.

It is not in the national interest to make sure that the Intel common stock remains valued above 0. Plenty of companies that are of national interest were saved directly from bankruptcy, by nationalizing them in a way that wipes out the common stock.

Depending on how this was done and for what quid pro quo, this could wind up being a liability for Intel. (It also trashes its image abroad.)