Why did Intel dominate, then lose?

The setup: DRAM, then microprocessor

Intel was founded in July 1968 by Robert Noyce and Gordon Moore, with Andy Grove as the third hire. The starting product was dynamic random-access memory — DRAM — which was a new technology in 1968 and which Intel had a manufacturing process edge in. For the first six years the company's identity was a memory company; the 1101 SRAM and the 1103 DRAM were its name-making products, and they competed against established memory makers like Mostek, Texas Instruments, and Inmos.

In 1971, almost as a side project, Intel designed a four-bit microprocessor — the 4004 — for a calculator maker called Busicom. The chip was a curiosity rather than a strategic bet. The 8008 followed for a separate customer (Datapoint, for a terminal), and then the 8080 (1974) was the first chip Intel sold openly into the broader market. The microprocessor business grew slowly through the 1970s while memory stayed the bigger line.

The shift came at the start of the 1980s when Japanese DRAM manufacturers — Hitachi, NEC, Toshiba, Fujitsu — began undercutting US prices on memory by 30-40%. By 1984 Intel was losing money on every DRAM it shipped. In a now-famous conversation, Grove asked Moore: 'If we were kicked out and the board brought in a new CEO, what would he do?' Moore answered: 'He would get us out of memories.' Grove said: 'Then why don't we walk out the door, come back, and do that ourselves?' They exited DRAM in 1985 and bet the company on microprocessors. The 386 (1985) and 486 (1989) confirmed the bet.

Source: Andy Grove, *Only the Paranoid Survive* (1996); Intel oral histories, Computer History Museum.

The Wintel deal

The IBM PC launched in August 1981 with an Intel 8088 processor and Microsoft DOS. IBM had picked the parts off the shelf in a hurry; the PC project had been an outsider effort inside IBM and the team had used commodity components to ship faster. The decision that mattered was IBM's choice to let Intel sell the same chip to PC clones, which Compaq, Dell, HP, and a hundred smaller companies started doing within three years. By 1990 the open-PC ecosystem was the dominant compute platform on earth, and Intel and Microsoft together collected most of its margin.

The Wintel arrangement was not a formal alliance. It was the result of binary architectural lock-in: each new generation of Windows assumed each new generation of x86 instructions, and each new generation of x86 was tuned for what Windows needed. Application developers wrote against the combined stack, and switching costs accumulated on every side. By 1995 Intel had gross margins above 65% on a commodity product, which is what happens when you have a hundred competing system integrators and effectively zero competing CPU suppliers.

For fifteen years this was the most lucrative business in technology. Intel built its own fabs, designed its own chips, packaged them in house, and sold them at premium prices to a market it had effectively cornered. Every two years Moore's Law gave it a process upgrade that competitors could not match. AMD, the only real alternative, was usually a generation behind on process and stayed in business as a cheap-fallback supplier for risk-averse buyers.

The Itanium misstep

Around 1994, Intel and HP started designing the architecture that would replace x86 — a 64-bit instruction set called Itanium, based on a then-fashionable design philosophy called EPIC (Explicitly Parallel Instruction Computing). The bet was that 64-bit computing was coming, x86 was a 1970s instruction set with too much legacy cruft to evolve cleanly, and a clean-sheet design would deliver a generational leap.

Itanium took longer to ship than planned. The first chip arrived in 2001, five years late, and was slower than the parallel-evolved x86 chips at every workload that mattered. The compilers needed to extract parallelism from Itanium code never matured. Customers refused to port. Meanwhile AMD, working with much less budget, designed AMD64 — a 64-bit extension of x86 that preserved every existing application — and shipped it in the Opteron in 2003. Microsoft picked AMD64 for the 64-bit version of Windows. Intel was forced to license AMD64 (under the cross-license agreement) and adopt it for its own 64-bit chips. Itanium limped along in niche server markets until Intel formally killed the product line in 2021.

The damage was deeper than the product failure. Intel had spent years and billions of dollars on the wrong architectural bet, while AMD — running on a fraction of the budget — had defined the 64-bit transition for the industry. Intel's reputation for architectural leadership took a serious hit, and the implicit guarantee that 'Intel always wins the next generation' was punctured for the first time.

The process collapse, 2015-2024

For thirty years Intel's competitive advantage had been process. Even when its designs were occasionally outshone by competitors' architectures, it manufactured them on a node ahead of everyone else, which usually compensated. The wheels came off this advantage starting around 2014-15.

Intel's 14nm node shipped on schedule in 2014. The 10nm node was supposed to ship in 2015 or 2016. It did not ship in volume until 2019 — four years late. Various technology reasons have been cited: an ambitious bet on cobalt interconnects, on a high-density library design, on certain self-aligned quad patterning techniques that were difficult to yield. The corporate-culture reasons are less debated: Intel had become organized around a process roadmap that assumed Intel would always lead, the lead had created complacency, and bad news from the fab floor took too long to escalate. When the 10nm program slipped, the company kept promising near-term delivery dates that kept getting missed.

While Intel was stuck on 14nm and its variants from 2014 to 2019, TSMC delivered 10nm (2016), 7nm (2018), 5nm (2020), and 4nm (2022). Apple, NVIDIA, AMD, Qualcomm, and the rest of the fabless ecosystem all migrated to TSMC. AMD in particular — which had been a generation behind Intel on its own internal manufacturing — pivoted to using TSMC nodes, and the resulting AMD chips (Ryzen, EPYC) suddenly had a process advantage rather than a process disadvantage. Intel went from leading by a generation to trailing by one, then two.

The 7nm node was meant to recover. It did not. In July 2020, Intel disclosed that 7nm was delayed by another year, and the stock dropped 16% in a day. By 2021, Intel's process technology was approximately two years behind TSMC, and the gap was not visibly closing.

The Gelsinger pivot and the IDM 2.0 bet

In February 2021, Pat Gelsinger — a 30-year Intel veteran who had left for VMware in 2009 — returned as CEO. He announced an aggressive plan called IDM 2.0: Intel would (1) keep designing chips, (2) keep its own fabs but also outsource selectively to TSMC for some products, (3) reopen its fabs to external customers as a foundry business, and (4) commit to five process nodes in four years to close the TSMC gap by 2025. The plan was ambitious to the point of being implausible. Industry observers gave it 30/70 odds at best.

By 2024 the plan was failing on the foundry leg. The Intel Foundry Services line was losing roughly $7 billion per year on $19 billion of revenue, and the external customer wins were minimal. Intel announced a 15,000-person layoff. The board removed Gelsinger in late 2024. The CHIPS Act money the US government had committed in 2022 — about $8 billion for Intel of the $52 billion overall package — was repurposed under the new administration to take an equity stake in the company. Intel was no longer a normal semiconductor company. It was a national champion the US government had begun underwriting.

The technical news was somewhat better. Intel's 18A process — the company's planned recovery node — appeared on schedule in 2024 production-qualification, and the first customer (Microsoft, for an AI accelerator) was named. Whether 18A volume actually closes the gap with TSMC's 2nm is the unresolved question of 2026.

Source: Intel 10-Q filings 2023-Q4 through 2024-Q3; Reuters and Bloomberg coverage of the 2024 Gelsinger removal; CHIPS Act amendment Q4 2024.

Strategic read

What Intel lost was not technology. It was the customer base. For three decades, the PC plus enterprise server market was so dominant that an integrated supplier could amortize the cost of staying on the leading edge. Once mobile became the bigger market (post-2010), and once the hyperscalers started designing their own server chips (post-2015), and once AI accelerators became their own pricing tier (post-2020), the customer base for x86-as-a-bundle compressed. The remaining x86 buyers were not paying enough margin to fund a $30 billion fab buildout every two years.

TSMC won by being the fab to the rest of the world. It does not design chips and therefore competes with no customer. Apple, NVIDIA, AMD, Qualcomm, Broadcom, Marvell, MediaTek, and a hundred fabless designers each placed their volume on TSMC nodes, and the combined volume amortized leading-edge fab cost in a way no single integrated company could. By 2025, TSMC was the most valuable semiconductor company in the world and Intel was a distant fourth or fifth depending on the day.

The interesting strategic question for AI infrastructure is whether vertical integration returns at the workload tier. NVIDIA does not own a fab, but it owns the software stack (CUDA), the system design (NVLink, NVL72), the supply contracts with HBM and packaging, and the customer relationships. Apple owns the OS, the chip, the package, and the device. Google owns the model, the TPU, and the datacenter. Each of these companies has rebuilt the integration that Intel used to embody — but using outsourced manufacturing for the one piece that requires $30B in capex per generation. The lesson of Intel is not that vertical integration is dead. It is that you should own the layers where the customer relationship lives, and outsource the layers where the capex base requires more volume than you can provide. Intel got that decomposition wrong; the modern winners got it right.