No, I am not going to begin talking about what happened yesterday with ASML. Before we get there, we need to start from the 11th of October when this article (of course totally ignored by MSM and Social Media busy broadcasting the latest phenomenal predictions Jensen Huang was spitting out non-stop) appeared on the web: $2 H100s: How the GPU Rental Bubble Burst.
This is the article TLDR: Don’t buy H100s. The market has flipped from shortage ($8/hr) to oversupplied ($2/hr), because of reserved compute resales, open model finetuning, and decline in new foundation model co’s. Rent instead
Let me summarize here the key points:
- At the very beginning of the AI race, building a data center filled with NVIDIA H100 GPUs and renting them out for at least $4.7 per hour allowed the providers to recover their investment in infrastructure in about 1.5 years.
- Overall through the years, $600bn USD of data center infrastructure investments have been poured to accommodate AI demand believed to be of a nearly infinite order of magnitude.
- In August 2024, several major providers were already slashing their rates for H100 GPU rentals to $1-2 per hour to attract demand that, apparently, wasn’t as infinite as people projected.
- While before service providers expected to effectively “print money” after 1.5 years, at the current prices they won’t be able to recover their investments in less than 5 years, and while the supply keeps increasing, the demand isn’t matching it.
- Many AI companies entered into multi-year contracts in 2023 to make sure they locked in the prices (at that time still rising) and GPU computing capacity, agreeing to rates above $4 per hour. However, these companies are now using a fraction of the capacity they bought due to the improvements in training and implementing AI models. As a consequence, in order to recover some of the costs they already agreed to pay, they became resellers in the market, pushing prices even lower.
- The current supply doesn’t even take into consideration the H100 GPUs datacenters coming online now after buyers finally got their GPUs delivered after many months of delay.
- Customers are now shifting to cheaper alternatives, like AMD, especially for inference models since the revenue stream they projected isn’t materializing as expected, and they are now under pressure to reduce their cash burn while increasing their efficiency and sustainability of their operations.
This article doesn’t even take into consideration the new Nvidia Blackwell that, according to Jensen Huang, was flying off the shelves:
- Nvidia rides fierce Blackwell demand to record stock close
- NVIDIA Blackwell GPUs Sold Out: Demand Surges, What’s Next?
- NVIDIA’s Entire Blackwell AI GPU Supply Sold Out For The Next 12 Months, AI Firms Gobble Up 100K Units In A Single Order Highlighting Insane Demand
- Nvidia CEO Jensen Huang says demand for next-generation Blackwell AI chip is ‘insane’
Clearly, the two stories did not match at all, and this is when ASML made the “mistake” of releasing its financial results earlier than scheduled. What did ASML accidentally reveal? In a nutshell, contrary to Wall Street analysts’ very bullish expectations for ~$5.6bn of net bookings, ASML disclosed $2.6bn or less than half of them, causing a 16% crash in the company stock that wiped out $50bn of value and cascaded on all other hyped semiconductor companies from Nvidia to Arm. There was one comment in particular that investors did not like from ASML’s CFO, who claimed the demand for their machines (that are used in the production of high-performance semiconductors like GPUs) has been weak because their customers like TSMC are currently running at ~80% capacity.
Hold on a second. Wasn’t Jensen Huang claiming “insane” demand for Blackwell, and how was all production for the following year already fully booked? If this is the case, how is it possible that TSMC (the only manufacturer of Nvidia products) isn’t operating its factories at full capacity? Let me play devil’s advocate here: is there a chance that ASML released the wrong financial statements? I mean, they released those including the real state of its operations rather than those meant to be provided to Wall Street after future orders were “adjusted” in order to be in line with the narrative so adamantly pushed by Nvidia? It is not a secret how companies often have last-minute “reviews” of their accounts before they are made public, with instances in the past when orders have been booked right before financial statement releases in order to embellish a quarter that otherwise would have been underwhelming. Personally speaking, it is hard not to suspect something of this sort could have been involved with ASML, considering the key role this specific company has in the whole semiconductor ecosystem.
Ironically, one day before the article I mentioned at the very beginning, I published THE SMOKING GUN THAT PROVES HOW OPENAI IS MICROSOFT’S REVENUES LAUNDROMAT demonstrating point blank how Microsoft is using a complex scheme of credits for the use of its Azure data centers to inflate its revenues and effectively hiding the oversupply problem that has been documented in detail by a tech analyst who aimed to give suggestions to other tech developers without realizing the broader implications of all the evidence he brought to light, especially for Nvidia (articles archive)
JustDario on X | JustDario on Instagram