Lip-Bu Tan took the hardest job in chips to save Intel

Sarah Guo and Elad Gil open No Priors with Lip-Bu Tan, the Walden investor turned Cadence CEO who now runs Intel. The obvious question is why anyone would take what he himself calls the hardest job in the industry. He is 66. People assumed he should retire. His answer is that Intel is an iconic company, too important to the semiconductor ecosystem and to the United States to walk away from. After Cadence, he wanted to do one more.

Asked what surprised him most in the past year, he does not name a product or a customer. One early morning, President Trump asked him to resign over a conflict of interest, with no exceptions. Tan first convinced himself of a simple line: he does not need this job; he is doing it purely to save Intel. With the personal issue set aside, the only remaining question was what he could still do to help the company.

The meeting was set for Thursday morning. By Monday he was in the room, and Trump listened. Tan walked through his biography: born in Malaysia, raised in Singapore, MIT, a life lived in the United States, never a long stretch in another country. Something in that account landed. Trump gave him the chance, and Tan says he is delighted to still have the work. Fourteen months in, the scoreboard he cares about is already in motion: culture, the balance sheet, and the product roadmap.

Saving Intel, in his telling, starts with culture. He is used to startups that move at the speed of light and has no patience for layer upon layer of meetings. He wants faster decisions, real accountability, and a habit of listening to customers. Customers, he says, are delighted that someone at Intel is humble enough to hear their problems and try to fix them. From day one he had every engineer report to him. He is an engineer by training. He wanted to know what went wrong before he tried to correct it.

The operating rhythm is the one he used at Cadence: crawl, then walk, then run. Crawling means being humble and listening. The first concrete step was the balance sheet, which he calls heavy and, in some ways, horrible. He explained to Trump that TSMC started with the Taiwan government as a shareholder, and that Japan and Singapore backed their industries the same way. Semiconductors are infrastructure. The U.S. government should provide support. It became a big Intel shareholder.

Old friends showed up too. Jensen Huang, a longtime friend, invested $5 billion. That stake is now worth $25 billion or more, which Tan reads as evidence that some of the work is landing. Masayoshi Son of SoftBank, where Tan once sat on the board, also lent a hand. With the balance sheet stronger, he simplified the product line, kept listening to customers, and set a five-to-ten-year roadmap for the next generation of leadership products. He does all recruiting himself, with no search firm, because a Rolodex of who to call still matters.

Luck, he admits, is part of the story. Agentic AI and inference have made CPUs highly demanded again. Training clusters used to mix CPUs and GPUs at one to eight. He now sees one to four, and maybe one to one. Talking to model companies and developers, he hears that reinforcement learning and the speed of orchestrating agents actually run better on CPUs. Demand for Intel processors, he says, is very high, and he is happy about that. The data-center server franchise is something to build on.

Foundry is the harder front. It is capital intensive, and when he took over the market was full of voices saying it was too expensive and would not work. He still doubled down, because any large semiconductor company needs a robust, resilient supply chain and cannot depend on one or two players in a few geographies. More customers, he believes, will conclude that making in the United States is critical. Foundry is a service business and a trust business. If yields fail, the customer's revenue misses, and they are toast.

So he watches yield, defect density, and cycle time, and he insists on the right IP, such as a low-power kit if you want to win mobile. Customers are already asking for more than silicon. Some want the whole rack, which means software and systems, a full stack he says he is quietly assembling. Then there is Elon Musk. Tan calls him one of the best entrepreneurs of the century, perhaps the best. They share the view that semiconductor infrastructure has not kept up with AI growth in capacity, productivity, and efficiency. Musk needs vast silicon for robots and cars, so he is building his own fab, Terafab, using Intel process and technology. The teams meet weekly. Musk questions every traditional step. Tan will not let anyone smoke in the clean room, but he will listen.

Elad Gil sketches the macro before Tan lists the bottlenecks. Claimed AI layoffs, he says, are mostly overstated; much of the cutting is a hangover from 2020 over-hiring, and the first heads to go are outsourced support and IT, which hits BPO-heavy countries such as the Philippines and India. Tan's frame is larger. AI's impact will be bigger and more profound than the internet. The first wave is efficiency: agents doing mundane work faster, and chip design itself coming out cheaper and sooner. The companies most at risk are the ones that do not embrace AI.

Three constraints get less airtime than they deserve. Power is the obvious one: some countries simply do not have the electricity, and AI expansion stops there. Helium, which many people do not realize, can hit semiconductor manufacturing hard. Memory is already in a scramble, and standing up new fab capacity takes years. CPUs and GPUs are tight too, so costs will be passed downstream and prices will rise. Labor cost is real, but it is not why he stayed in foundry. He stayed because the industry just lived through supply-chain shock, and the most advanced processes only get harder.

Intel's 18A is in place; 14A, about 1.4 nanometers, is going to production; he is already planning 1 nanometer and 0.7. The features are hair-thin. One mistake and the lot goes down the drain. Moore's Law still wants performance to double without doubling cost and area, which gets more expensive unless materials and design change. He is hiring materials scientists and has invested across gallium nitride, silicon carbide, and indium phosphide. For packaging he is betting on Intel's EMIB-T against TSMC's CoWoS, on glass substrates via 3D Glass Solutions, on artificial diamond through Diamond Foundry, and on a large advanced-packaging program with India and New Mexico. Engineers, he says, always hit a wall. Then they jump over it or go around it.

Eighteen years ago he would walk into a tier-one venture partnership to pitch semiconductors and watch half the room invent excuses to leave. The rest asked if he had a software deal. Two people stayed out of sympathy. The tape has reversed. Nvidia, in his rundown, is a $5.3 trillion company; Broadcom and TSMC sit around $2 trillion; Lisa Su's AMD is almost $800 billion; Intel is close to $600 billion. VCs who once sat out now want in. He has 159 IPOs and 126 acquisitions on the board, including more than 200 semiconductor investments, 38 percent of them in the United States.

The method is to start with the bottleneck. Interconnect got tight, so he backed names such as Credo and Astera Labs, plus Celestial AI on the optical side; Jensen Huang, he notes, has been putting money into almost every photonics-related company. Power conversion from 40 volts down to 1 volt wastes energy, so he backed integrated voltage regulation at Empower, later bought by Analog Devices. Inphi, an optical interconnect chip company, was acquired by Marvell. New EDA that uses machine learning to cut design cost looks like a gold mine: Cadence under his successor Anirudh Devgan has gone all-in on AI, Synopsys took a $2 billion Nvidia investment and bought Ansys to reach system-level design, and startups can still go public or sell to Cadence, Synopsys, or Siemens.

From day one, target a first customer, preferably a hyperscaler that will spend millions over a few years; a warrant is worth it if that logo lets you scale. Talent he hunts in Silicon Valley, Austin, and Israel, where he has put a significant share of his checks. Israeli founders, even in wartime, take the conference call, drop into a shelter when the warning sounds, and switch to voice if the video dies. That resilience, he says, is the entrepreneurship he enjoys. Nine of ten companies he funds change their business plan halfway because the market moved, so he backs a team, not a lone hero: open, willing to hear coaching, and still able to draw their own conclusion.

Guo asks how industrial policy should sit in American business culture, which long frowned on it even as TSMC became the most important company in the world. For capital-intensive infrastructure, Tan says you need government funding, sovereign wealth, and the big infrastructure funds. Venture itself has changed: billion-dollar checks, once unheard of, now happen, and Series A valuations already clear $1 billion, so the 10x has to be caught earlier. Mutual funds are coming into pre-IPO rounds and care less about owning 20 percent, which is lucky, because there is not that much 20 percent to give. As a public-company CEO he wants long-term growth holders, not the ones whose first question is the buyback.

What do investors most misunderstand about Intel right now? On foundry, he does not pretend. The company is distant from TSMC's performance. The work is to be humble and stack the building blocks he keeps naming: IP, yield, defect density, cycle time. Customers have to trust you before they hand over wafers. That path is slow. He expects the business to start surfacing around 2030 to 2032.

On products, the misread is deeper. PCs remain the cash-flow base, but that is not the whole story. He is quietly assembling CPU, GPU, and software architects so Intel can leapfrog, and he wants the company to run at a multiple of startup culture. Training, he grants, is Jensen's. The jumble of edge, agentic AI, and physical AI is still open. The old job was serving humans with PCs and servers. The new dimension is millions of agents that need compute and a way into the software stack. Put XPU, advanced packaging, and foundry together and you can build purpose-built silicon for those workloads. Fourteen months produced a 6x for shareholders. Being a venture capitalist at heart, he is hunting 10x over five to ten years. Cadence was 76x as CEO and about 85x as executive chairman from $2.42. Intel's base is too large to repeat that. Ten times, he says, would be a good return.

Guo puts a live industry split on the table. One camp says compute keeps concentrating: a gigawatt data center is the beginning, and even inference is more efficient in the middle. The other camp thinks edge and client will take a large share. Tan agrees the massive AI infrastructure build is the right move. Workloads are still rising. The only thing that slows it is supply constraint. After the buildings go up, though, you have to come back to the application. Not everyone who builds will win.

He has watched this movie. In the internet era, Amazon and Netflix became the application winners; others went sideways, disappeared, or were acquired. AI will grow, then consolidate, and leave one or two true winners in a crowded lane. If a market is too packed, maybe one or two survive and the rest merge. His own filter is whether the application is huge, sustainable, and not already a stampede. Elad adds the physical-world version from robotics and defense: what you assume about on-device compute and home connectivity decides what a robot can do, a question the SaaS era forgot and that is now back.

Tan's investment thesis matches that. Find a problem that really needs solving, find players you can partner with, judge how big and durable the application is, and if you believe it, double or triple down, even before the application is broadly deployed. Inside Intel the same crawl-walk-run applies to people. He first hired the best semiconductor talent, then software people for a full stack, then noticed the team's average age sitting in the late forties and fifties. He needs people who understand frontier models and open source. The teacher, it turns out, is at home: when he visits his son and grandchildren, he taps him on AI and machine learning. Intel, once a legacy spreadsheet company, is being turned into an AI-enabled one, combining veterans, new blood, and the best tools he can put in design, sales, and marketing.