Apple Silicon After the CPU Era.Original analysisNot investment advice

For most of the last decade, the Apple Silicon story was simple. Each year, the new iPhone shipped with a chip that was meaningfully faster on the CPU benchmarks people argued about on the internet. Single-thread leadership was a marketing pillar. Then in 2021, the A15 broke the rhythm. The CPU was excellent but the year-on-year jump looked unusually small. A widely read piece argued that the brakes were already on, that Apple had moved silicon budget away from CPU cores, and that the engineers behind the great years had left for Nuvia and Rivos. Five years later, the verdict is more interesting than either “Apple is fine” or “Apple is finished”.

The 2021 piece was directionally right that easy yearly CPU leaps were ending. It was incomplete in concluding that Apple’s silicon edge would erode in step. What actually happened is that Apple stopped competing primarily on CPU and started competing on the entire chip: GPU, Neural Engine, image and media engines, memory bandwidth, cache, thermals, and the operating system that knows how to use all of them. The CPU kept improving. It just stopped being the whole story.

Apple’s CPU gains did not die. They became less spectacular, less isolated, and less important than the system around them.

---

1. Why A15 looked like a warning sign

The setup, in 2021, was that Apple had trained the market to expect obvious year-on-year CPU jumps. The A15 still shipped a strong chip, but the CPU uplift looked unusually small relative to prior Apple cadence. The uploaded SemiAnalysis article framed the move as a redirection of silicon budget away from CPU cores and toward the rest of the system.¹

The argument was anchored in a small number of structural changes that were not, in themselves, controversial.

You can read those changes two ways. The bearish reading is the one the article emphasised: a thinner architectural CPU change, a benchmark uplift close to the clock-rate delta, and a transistor budget pointed at GPU, ISP, media, display, and neural blocks instead of a major CPU redesign. The neutral reading is that 2021 was the first year Apple openly traded CPU front-page numbers for a fatter system-level chip. The piece was directionally right in calling the slowdown. It is the second reading that ages well.

It is also worth being careful with the benchmarks themselves. The 7.7% / 9.3% figures came from early Geekbench-class samples. User-submitted benchmark averages are useful for trend, not for surgical claims about microarchitecture.¹¹

---

2. What actually happened after A15

If A15 was the year the cadence broke, A17 Pro through A19 Pro is the period that resolves the argument. The CPU kept gaining, but the loudest year-on-year numbers Apple now puts on stage are no longer about the CPU.

Two patterns inside that timeline matter more than any single chip.

First, the GPU and Neural Engine became the loud parts of the story. A17 Pro headlined a GPU redesign and ray tracing. A18 headlined Apple Intelligence with a much faster Neural Engine. A19 Pro put neural accelerators in every GPU core for local LLM workloads. The CPU still gets percentage gains in Apple’s press copy, but it is no longer the protagonist.

Second, the comparisons stretched out. Apple began comparing A19 to A15, not to last year’s part. The honest read is that single-year, single-thread CPU gains compress as the field matures. Multi-year gains still compound. Both can be true.

The CPU kept improving. But it stopped being the whole story.

---

3. The battlefield moved from CPU to system silicon

The reason the 2021 framing aged unevenly is that Apple did not stop investing. It re-allocated. A modern Apple chip is a coordinated system, not a CPU with peripherals. The CPU is one block among several, and increasingly not the one that determines whether a feature is plausible on-device.

This is not a story unique to Apple. The whole industry has shifted toward accelerator-heavy SoCs. Apple just had a head start, vertical integration across hardware and OS, and a willingness to spend silicon on blocks that do not show up cleanly on a CPU benchmark. The CPU era was about IPC and clock speeds. The system silicon era is about which workloads you can run on device without phoning the cloud.

Looked at this way, the 2021 article was correct that Apple was redirecting silicon budget. It was wrong only if you assumed CPU front-page numbers were the only currency Apple cared about. The currency had already changed.

---

4. The manufacturing backdrop

None of this happened in isolation. The reason every leading-edge chip designer has shifted toward accelerator-heavy systems is that the foundry side of the industry has changed underneath them. Easy node shrinks are gone. Process improvements still exist, but each step is harder, more expensive, and less automatic.

The shorthand for that shift, on the leading edge, is three letters at a time. EUV stands for extreme ultraviolet lithography, the wavelength tooling that made the 5nm and 3nm generations economical. High-NA is the next generation of EUV with a wider lens, currently being matured for high-volume manufacturing. Nanosheet transistors replace FinFETs at 2nm and below. 3D integration, which includes stacking and advanced packaging, is the industry’s answer to the limits of pure 2D scaling.

The contours are public. TSMC says its N2 family entered volume production in the fourth quarter of 2025 and uses first-generation nanosheet transistor technology.⁷ Its 3nm family covers N3, N3E and N3P with incremental power, performance and density improvements.⁸ ASML’s 2026 AGM framing is the part that matters for this essay: Moore’s Law alone is not sufficient to meet future AI training compute needs, and the future is framed as 2D scaling plus 3D integration.⁹ ASML’s Q4 FY2025 deck adds the supply-side line: AI demand for compute density and high-bandwidth memory is growing faster than overall semiconductor sales, EUV capacity has grown strongly, the NXE:3800E has reached 230 wafers per hour, and High-NA systems are being matured for high-volume manufacturing requirements by end of 2026 with insertion in 2027 to 2028.¹⁰

None of those numbers prove anything specific about Apple’s chip, and you should not let them. What they do explain is the climate. In a world where 2D scaling is harder and more expensive, the chip designer who allocates budget across the full system gains relative to the chip designer who keeps pouring silicon into the same CPU core. Apple was early to that re-allocation. The 2021 piece called it a CPU slowdown. The honest 2026 reframing is that the industry it was set inside had moved.

---

5. The Nuvia and Rivos warning aged better than expected

The piece’s second argument was about talent. Apple’s star CPU engineers left for Nuvia and Rivos. The piece treated those departures as a CPU roadmap risk for Apple. That part is more interesting than it first looked.

Qualcomm completed its acquisition of Nuvia for around USD 1.4B in 2021.¹² The Nuvia CPU work became central to Qualcomm Oryon, which Qualcomm now positions as the core in the Snapdragon X Series for PCs and the Snapdragon 8 Elite for phones.¹³ Rivos became part of the RISC-V AI infrastructure story. Reuters reported Rivos raising about USD 250M to develop RISC-V AI chips,¹⁴ and later reported that Meta planned to buy Rivos to strengthen its internal semiconductor work.¹⁵

What this adds up to is not that Apple collapsed. It is that the brain drain helped seed a generation of custom-silicon competitors outside Apple. Qualcomm now has a credible PC CPU thanks to ex-Apple-and-Nuvia talent. Meta is buying an AI-chip company built by ex-Apple engineers. The 2021 piece treated the exodus as Apple’s problem. With distance, it looks more like a problem for everyone who used to compete with Apple without one.

The brain drain did not destroy Apple. It did help seed the next generation of custom silicon competitors.

---

6. Evidence ledger

Because this essay sits between a 2021 thesis and a 2026 reading, the source layer matters. The ledger below pairs each load-bearing claim with a primary source and a one-line note on why it matters.

---

7. The verdict

The original A15 warning was right about the end of easy annual CPU leaps. It was wrong only if read as “Apple is finished”. Apple’s advantage did not disappear. It moved. The modern Apple chip is a coordinated system of CPU, GPU, Neural Engine, ISP, media engines, memory, cache, thermals, software frameworks, and product-level integration. Each of those blocks gains less attention than a clean “X% faster” CPU number. Together, they decide which features ship on device at all.

The Nuvia and Rivos warning aged better than its CPU framing alone. The brain drain mattered, not because Apple collapsed, but because ex-Apple talent helped seed credible custom-silicon competition outside Apple. Qualcomm has a real PC CPU. Meta is buying an AI chip company. That is not the failure of Apple Silicon. It is the success of the playbook that built it.

The honest 2026 statement is short.

Apple’s CPU gains did not die. They became less spectacular, less isolated, and less important than the system around them.

---

Sources & footnotes

This essay treats the uploaded 2021 SemiAnalysis article as a cited historical anchor. Specific claims about modern Apple chips are sourced from Apple Newsroom press releases. Industry framing is sourced from TSMC technology pages and ASML investor and AGM material. Competitor and talent claims are sourced from Qualcomm investor materials and Reuters reporting. No third-party charts, screenshots, or logos are reproduced. This is not investment advice.