Buzz about Broadcom’s custom chips is testing, but not breaking Nvidia’s dominance
If money were no factor, custom chips like Broadcom’s would always be the default choice for building cutting-edge AI models, says a Google software engineer. “If resources are not limited and I need to move quickly and get as much training as possible, I use a 100 percent custom chip,” said Gabriel Rasskin, who works on the Gemini AI team. “Every second of computing counts,” he emphasized in an interview with CNBC. Led by Broadcom, this mentality also appears to be growing among hyperscalers as the appetite for custom chips grows. Google successfully trained Gemini 3 using tensor processor units (TPUs) co-designed by Broadcom. Gemini 3 put the Alphabet unit back on the map as the big language model to beat with its November launch. The performance of TPUs also brings Google into the conversation as an alternative AI chip to industry-standard Nvidia GPUs, or graphics processing units. Every second of calculation counts. Gabriel Rasskin, Google Gemini software engineer Specialized chips are designed for specific high-volume tasks; Nvidia’s general-purpose GPUs are not designed to do this. But Nvidia CEO Jensen Huang recently dismissed that, saying custom chips were a threat to his business. “What Nvidia is doing is much more versatile,” he told Jim Cramer in an interview last month. “Nvidia can appeal to much broader markets, not just chatbots.” In November, Nvidia said in a post about X that it was “delighted with Google’s success.” 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 wherever computing happens.” Of course, other competing custom chips, also called TPUs or application-specific integrated circuits (ASICs), won’t replace Nvidia. Google is also one of Nvidia’s biggest customers; It relies on multiple GPUs to power its cloud infrastructure and offer customers the computing power to train and run their own AI models. Google, which has been developing TPUs for more than a decade, is also offering their use to cloud customers. Google Cloud is the third largest cloud, behind Amazon and Microsoft, which also offer their own custom chips while remaining major Nvidia customers. While industry experts largely agree that Nvidia’s risk is minimal at this point, some believe Nvidia’s dominance in the AI chip market is being tested. “In any market, you’re going to lose some market share,” DA Davidson analyst Gil Luria told CNBC. “This is a free market, and profit attracts competition. That’s what’s happening here. Nvidia’s customers, especially the biggest ones, Amazon, Google, Microsoft, Meta, don’t want to be dependent on a single vendor,” Luria said. Broadcom’s custom chip deal with OpenAI, the company behind ChatGPT, in October underscores the diversification trend. Despite Broadcom’s interest, Luria said GPUs are still necessary. “Even for Google, TPU is just an add-on. In the past, they’ve mostly used it for their own internal purposes. Now they’re starting to outsource it, even selling it. But they’re still mostly buying Nvidia.” An Apple source familiar with the intricacies of chip development noted that the high barrier to ASIC production works in Nvidia’s favor. It can take too long and be too costly for small companies to track customer chips. Limited access to chip manufacturers’ production capacity is another constraint. Remember, Broadcom and Nvidia design the chips, but they’re considered foolproof, which means they actually turn to factories like Taiwan Semiconductor (TSMC) to make the hardware. Apple has been developing its own chips since 2010 and is slowly replacing third-party silicon in its devices. “Broadcom does the best in the custom chip category,” Luria said. In last month’s earnings call, Broadcom CEO Hock Tan said the company’s AI revenue “rose 65% year over year to $20 billion, pushing the company’s semiconductor revenue to a record $37 billion for the year.” The CEO also confirmed the rumors we heard earlier that the fourth customer, who placed a $10 billion order, was indeed Anthropic. “But unlike Nvidia, which you can reasonably predict will remain a dominant force, Broadcom’s position is more fragile,” Luria said. he said. “Its biggest customer so far is Google. If Google decides to go directly to TSMC, as Apple did, that would present a much greater risk to Broadcom than the risks facing Nvidia.” Luria said he expects Nvidia to maintain more than 50 percent of its market share for at least the next five years and more than 70 percent for the next three years. Analyst DA Davidson has a buy rating on Nvidia with a price target of $250 per share. This represents an increase of almost 31% from Thursday’s close. Others on Wall Street are watching closely. Morgan Stanley this week reissued its buy-equivalent rating on both Broadcom and Nvidia, noting that “Nvidia remains preferred despite growing ASIC enthusiasm.” Analysts have cited what they call the industry’s highest ROI solution in cloud computing, with the Vera Rubin platform set to take off in the second half of 2026. But Wolfe analysts took a slightly more optimistic tone on Broadcom. “Google’s willingness to make TPU available to third parties creates a true competitor for Nvidia,” analysts wrote, adding that they expected Broadcom to be the primary beneficiary. The firm, which upgraded Broadcom to a buy-equivalent rating on Friday, forecasts shipments of about 7 million TPUs by 2028. Analysts issued a $400 price target for Broadcom’s shares; This represents an increase of approximately 21% from Thursday’s close. On Friday’s “Squawk on the Street,” Jim Cramer called the upgrade a “timely tech tip” for investors looking for a quality name that’s down nearly 20% from its record close of about $413 and 4% year to date. “I don’t understand the bad action at Broadcom,” Jim said during the CNBC Investment Club’s January Monthly Meeting. Jim’s tendency to buy the dip reflects the stance he took when Broadcom tumbled 11.4% in December after announcing a strong quarterly rise that was marred by misinterpreted comments. AVGO NVDA 5Y mountain 5-year performance of Nvidia and Broadcom Although Nvidia’s shares are slightly higher on a yearly basis, it has its own problems. Stocks are struggling with multiple pressure as investors become less willing to pay a premium for each dollar of earnings. The stock is also under pressure from ongoing US-China geopolitical tensions. Reuters reported this week that China has approved ByteDance, Alibaba and Tencent’s purchase of Nvidia’s H200 chips. Jim reiterated his “own it, don’t trade it” policy on the stock, but cautioned to be patient until deals are completed. According to Jim, Jensen’s presentation of the new Vera Rubin chips at Nvidia’s annual GTC conference in mid-March could help move shares in a positive direction. The club has an equivalent rating of 2 on both Broadcom and Nvidia, meaning we would recommend adding positions on significant pullbacks. Our Broadcom price target is $425 and Nvidia PT is $230. (Jim Cramer’s Charitable Trust is long GOOGL, AVGO, NVDA, AMZN, MSFT. See here for a full list of stocks.) When you subscribe 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 trading alert before buying or selling a stock in his charitable foundation’s portfolio. If Jim talked about a stock on CNBC TV, he would wait 72 hours after issuing the trading alert before executing the trade. THE ABOVE INVESTMENT CLUB INFORMATION IS SUBJECT TO OUR TERMS AND CONDITIONS AND PRIVACY POLICY, TOGETHER WITH THE DISCLAIMERS. NO CIVIL OBLIGATIONS OR DUTIES EXIST OR SHALL BE RESULTING FROM YOUR RECEIVING ANY INFORMATION PROVIDED IN CONNECTION WITH THE INVESTMENT CLUB. NO SPECIFIC RESULT OR PROFIT CAN BE GUARANTEED.
