If money were not a factor, a Google software engineer says custom chips, like those from Broadcom , would always be his default selection to build cutting-edge artificial intelligence models. “I am a hundred percent using a custom chip of some sort if resources are not a constraint and I need to move fast and get as much training as possible,” said Gabriel Rasskin, who works on the Gemini AI team. “Every second of compute matters,” he stressed in an interview with CNBC. That mindset appears to be growing among hyperscalers as their appetite for custom chips rises, with Broadcom leading the way. Google used tensor processor units (TPUs), co-designed by Broadcom, to successfully train Gemini 3, whose November launch put the Alphabet unit back on the map as the large language model to beat. The performance of the TPUs also put Google in the conversation as an AI chip alternative to the industry-standard Nvidia GPUs, or graphics processing units. Every second of compute matters. Google Gemini software engineer Gabriel Rasskin Custom chips are designed for specific high-volume tasks, something Nvidia’s general-purpose GPUs are not built to do. Nvidia CEO Jensen Huang, however, recently dismissed custom chips as a threat to his business. “What Nvidia does is much more versatile,” he told Jim Cramer in an interview last month . “Nvidia can address markets that are much, much broader, not just chatbots.” Back in November , Nvidia said it was “delighted by Google’s success” in a post on X. But the post also said, “Nvidia is a generation ahead of the industry — it’s the only platform that runs every AI model and does it everywhere computing is done.” To be sure, TPUs or any other competing custom chips — also called application-specific integrated circuits (ASICs) — are not going replace Nvidia. Google is also one of Nvidia’s biggest customers, relying on a massive number of GPUs to support its cloud infrastructure and offer customers computing power to train and run their own AI models. Google, which has been developing TPUs for more than a decade, also offers their use to cloud customers. Google Cloud is the third largest, behind Amazon and Microsoft, which have also been introducing their own custom chips while remaining big Nvidia customers. While industry experts largely agree the risk to Nvidia at this juncture is minimal, some believe Nvidia’s grip on the AI chip market is being tested. “In any market, you’re going to have some market share loss,” D.A. Davidson analyst Gil Luria told CNBC. “This is the free market, and profits attract competition. That’s what’s happening here. Nvidia’s customers, especially the largest ones, Amazon, Google, Microsoft, Meta , don’t want to be beholden to one vendor,” Luria said. Broadcom’s October custom chip deal with OpenAI, the company behind ChatGPT, underscores the diversification trend. Despite Broadcom’s traction, Luria said GPUs are still essential. “Even for Google, a TPU is just a supplement. They used it historically mostly for their internal purposes. Now they’re starting to also provide it externally, and they’re even starting to sell it. But they’re still mostly buying Nvidia.” A source at Apple , who is familiar with the intricacies of chip development, pointed to the high barrier for ASIC production, which works in Nvidia’s favor. It might take too long and cost too much for smaller companies to pursue customer chips. Limited access to manufacturing capacity at the chip fabricators is another constraint. Remember, Broadcom and Nvidia design the chips but are considered fabless, meaning they turn to fabs like Taiwan Semiconductor (TSMC) to actually make the hardware. Apple has been developing its own chips since 2010 and is slowly replacing third-party silicon in its devices. “Broadcom is doing the best in the custom chip category,” Luria said. On last month’s earnings call, Broadcom CEO Hock Tan said the company’s AI revenue rose “65% year-over-year to $20 billion, driving the semiconductor revenue for this company to a record $37 billion for the year.” The CEO also confirmed the rumors that the fourth customer we heard about on the prior, which placed a $10 billion order, is indeed Anthropic. “But unlike Nvidia, where you can reasonably extrapolate that it will remain a dominant force, Broadcom’s position is more fragile,” Luria explained. “Its biggest customer by far is Google. If Google ever decided to go directly to TSMC the way Apple did, that would represent a much bigger risk for Broadcom than anything Nvidia is facing.” Luria said he expects Nvidia to maintain over 50% of market share for at least the next five years and over 70% share for the next three years. The D.A. Davidson analyst has a buy rating on Nvidia with a $250 per share price target. That represents nearly 31% upside to Thursday’s close. Others on Wall Street are also watching closely. Morgan Stanley this week reissued a buy-equivalent rating on both Broadcom and Nvidia, while noting a “preference for Nvidia despite rising ASIC enthusiasm.” The analysts cited what they called the industry’s highest return-on-investment solution in cloud computing as the Vera Rubin platform ramps in the second half of 2026. Wolfe analysts, however, struck a slightly more bullish tone on Broadcom. “Google’s willingness to make TPU available to third parties creates a true competitor to Nvidia,” the analysts wrote, adding they expect Broadcom to be the primary beneficiary. The firm, which upgraded Broadcom to a buy-equivalent rating on Friday, estimates roughly 7 million TPU shipments by 2028. Analysts issued a $400 price target on Broadcom’s stock — that implies nearly 21% upside to Thursday’s close. On “Squawk on the Street” on Friday, Jim Cramer called the upgrade a “timely tech recommendation” for investors in search of a quality name that is down nearly 20% from its record close of nearly $413, and down 4% year to date. “I don’t get the poor action in Broadcom,” Jim said during the CNBC Investing Club’s January Monthly Meeting. Jim’s inclination to buy the dip echoes the stance he took when Broadcom sank 11.4% in December after posting a strong quarter beat that was clouded by misinterpreted commentary. AVGO NVDA 5Y mountain Nvidia vs. Broadcom performance 5 years Nvidia’s stock, while modestly higher for the year, has its own problems. Shares are battling multiple compression as investors grow less willing to pay a premium for each dollar of earnings. The stock has also been pressured by ongoing U.S.-China geopolitical tensions. Reuters reported this week that China approved ByteDance, Alibaba , and Tencent to buy Nvidia’s H200 chips. Jim reiterated his “own it, don’t trade it” policy on the stock, but cautioned patience until the deals are done. According to Jim, Jensen’s presentation of the new Vera Rubin chips at Nvidia’s annual GTC conference in mid-March could help move the stock in a positive direction. The Club has a hold-equivalent 2 rating on both Broadcom and Nvidia, meaning we would recommend adding the positions on significant pullbacks. We have a Broadcom price target of $425 and an Nvidia PT of $230. (Jim Cramer’s Charitable Trust is long GOOGL, AVGO, NVDA, AMZN, MSFT. See here for a full list of the stocks.) As a subscriber to the CNBC Investing Club with Jim Cramer, you will receive a trade alert before Jim makes a trade. Jim waits 45 minutes after sending a trade alert before buying or selling a stock in his charitable trust’s portfolio. If Jim has talked about a stock on CNBC TV, he waits 72 hours after issuing the trade alert before executing the trade. THE ABOVE INVESTING CLUB INFORMATION IS SUBJECT TO OUR TERMS AND CONDITIONS AND PRIVACY POLICY , TOGETHER WITH OUR DISCLAIMER . NO FIDUCIARY OBLIGATION OR DUTY EXISTS, OR IS CREATED, BY VIRTUE OF YOUR RECEIPT OF ANY INFORMATION PROVIDED IN CONNECTION WITH THE INVESTING CLUB. NO SPECIFIC OUTCOME OR PROFIT IS GUARANTEED.
Forrás: eredeti cikk
