The $3 Million Toaster ProblemBy Joel Litman, chief investment officer, Altimetry
 It was supposed to be the most advanced computer in the world...
Instead, the engineers were staring at a 3,000-pound, $3 million toaster. In March 2024, Nvidia (NVDA) unveiled its latest chip: the Blackwell B200. At a launch event – dubbed the "AI Woodstock" – CEO Jensen Huang promised the world that the Blackwell was up to 30 times faster than its predecessor. Each of these graphics processing units ("GPU") cost roughly $40,000. They were built as the engine of the next industrial revolution. Prior chips let AI models generate replies in seconds. Blackwell cut this down to milliseconds. For the biggest players in AI, this wasn't just an upgrade... It was a necessity. However, Nvidia's quest to get these chips to market had its fair share of hiccups along the way. In August that year, Nvidia admitted to a design flaw in the Blackwell that forced a multi-month delay...The chip's new two-part design was more efficient, but also more fragile. It couldn't handle the high temperatures generated during use. The chips warped and broke in the heat. After solving this issue, Nvidia began sending its first Blackwell GPUs to data centers like Microsoft's flagship facility in Phoenix, Arizona. That's where a new and far more expensive flaw was revealed...
| Recommended Links: |
|
 New System Predicts Stock Moves Using Satellite-Tracking Technology It could be the discovery of the century – and it could make you a LOT of money. For the past several months, a former algorithm expert from Goldman Sachs has been testing a theory... that stocks and satellites move in the exact same pattern over time. And he's launching a new stock-picking breakthrough that proves that theory to be correct. It just delivered gains of 263%, 444%, 678%, and even 2,487% in back testing... along with 100% win rates on many stocks. You can try it for free for a limited time. |
|
 SpaceX's 'Dark Energy' Could Replace Foreign Oil For years, we've been told that SpaceX is a rocket company. But according to new satellite images from 300 miles above the Earth's surface, there is something very strange going on at SpaceX right now that has nothing to do with space. It could soon replace our need for foreign oil forever and ignite a $10 trillion boom for the stocks involved. Click here to learn more. |
|
|
The setback wasn't due to any faults in the Blackwell itself. It came from the "plumbing" that powered it. To achieve its world-record speeds, a single Blackwell chip requires a massive surge of energy... roughly 1,200 watts ("W"). But this power needs to be delivered delicately. If energy is delivered at too high of a voltage, the GPU disintegrates. Instead, GPUs are fed very low voltages – around 0.8 volts ("V") for Blackwell chips. That's about half of what you get from a AA battery. But the electricity is flowing at a staggering rate – 1,500 amps of current. For context, a home typically handles around 200 amps. And that's through thousands of square feet. So the power flows 7.5 times faster than energy through a house, but it's going through something thousands of times smaller. Nvidia needed a way to push 1,500 amps through a device the size of a postage stamp...So it hired Monolithic Power Systems (MPWR) to figure out this "plumbing job." Monolithic specializes in voltage regulator modules. These devices act like transformers that regulate power going into your home. They convert large flows of energy into lower-voltage currents that are safe for chips to handle. The company was the leading module provider for Nvidia's prior GPU, the Hopper H100. And at $50 per unit, it was a cheaper choice for Nvidia than some competitors. Despite their low cost, these components were the only thing standing between 1,500 amps of raw energy and the most expensive silicon on Earth. If they failed, the energy wouldn't fuel the chip... it would destroy it. Both companies quickly learned that Blackwells were more difficult to maintain than Hoppers. Blackwell GPUs needed nearly twice as many amps as Hoppers did. And that means they produced nearly four times as much heat. Monolithic's regulators couldn't keep up. By November 2024, reports confirmed that the gatekeepers were failing...At the high-wattage levels Blackwell required, Monolithic's components burnt out – rendering entire 72-chip racks inoperable. Nvidia effectively fired Monolithic Power Systems for its top-tier Blackwell models as a result. The company canceled Monolithic's backlog and pivoted to competitors to save the launch. On November 11, 2024, MPWR stock plummeted 15%... wiping out $10 billion in market value. For Nvidia's customers, the situation was just as dire. Microsoft's Phoenix data center, which was designed to hold 50,000 Blackwell chips, sat partially empty. OpenAI reportedly became so concerned by the overheating "toasters" that it asked Microsoft for older, slower chips just to keep its AI training on schedule. Nvidia eventually resolved the 'toaster' crisis...It redesigned the racks to include specialized liquid cooling. But the incident brought an important reality to the forefront of the AI industry... A GPU is only as good as a company's ability to power it safely. Computing capacity gets the headlines. But power delivery is becoming just as critical to the AI build-out. As chips get faster, they also get more demanding. Blackwell needed nearly double the power of its predecessor. The next generation of chips will likely need even more. That puts pressure on the companies that convert and regulate that power. Miss the mark... and a $40,000 chip is worthless. Going forward, the winners in AI may not just be the fastest chipmakers. They'll also be the companies that can keep those chips running. Regards, Joel Litman
July 13, 2026
|
No comments:
Post a Comment