Why did Apple build its own chips?
Apple spent fifteen years on PowerPC, fifteen on Intel, and is now in year five of Apple Silicon. Each transition was driven by the same constraint — vendor performance per watt was no longer keeping up with what Apple wanted to ship — and each transition taught Apple's competitors that vertical integration at the silicon level is now within reach for any company with enough volume and enough patience.
Once your product volume is large enough to amortize a chip design team plus a foundry contract, doing your own silicon is the highest-leverage move available. Microsoft, Amazon, Google, and Meta have all reached that threshold for AI workloads, and they have all done what Apple did — for the same structural reasons.
The PowerPC era and how it ended
From 1994 to 2005, Apple shipped Macintoshes built around the PowerPC processor — a joint design from an Apple-IBM-Motorola alliance called AIM. The PowerPC architecture was good. The execution was not. Motorola's chip division was slow, IBM's was distracted by its server business, and Apple was the smallest of the three partners and got the least attention. Every CPU upgrade Apple shipped was late, hot, or both.
The breaking point came in 2003-04, with a chip called the PowerPC G5. The desktop version, designed by IBM, ran fast but ran hot — too hot for a laptop. Apple committed publicly to shipping a PowerBook G5, then spent two years watching IBM fail to produce a chip that could fit in the form factor at acceptable thermals. In 2005, Steve Jobs announced that Apple would migrate to Intel processors, abandoning PowerPC entirely. The transition completed in 2006. The PowerBook G5 never shipped.
The story is sometimes told as Apple choosing Intel. It is more accurate to say Apple was forced off PowerPC because its supplier could not deliver. The lesson — which Apple internalized completely — was that depending on someone else's roadmap for the most important component in your product is a strategic vulnerability that recurs whenever the supplier's priorities and yours diverge.
iPhone and the start of the silicon team
The original iPhone (2007) used an off-the-shelf Samsung ARM-based SoC. By 2008 Apple had decided that running iOS on someone else's chip — even one Apple co-specified — was the same trap as PowerPC, just at a smaller scale. The volume was already enormous. The chip architecture was going to be one of the few real differentiators in a category Apple expected to dominate. So Apple did two things.
In April 2008, Apple acquired PA Semi, a 150-person chip-design company in Santa Clara that had been working on low-power PowerPC processors. The acquisition was reported at about $278 million. In 2010, Apple acquired Intrinsity, a smaller firm that had built the high-performance Cortex-A8 implementation Apple was already using. Both acquisitions were primarily talent — Apple kept the engineers, redirected them to its own roadmap, and folded their work into a new internal silicon group led by Johny Srouji.
The first chip from this group was the A4, which shipped in the original iPad in April 2010 and the iPhone 4 in June. Externally it looked like a standard ARM-based SoC. Internally it was the first product of an in-house design team. Every iPhone and iPad since then has used an Apple-designed processor. The A-series chips became one of the most aggressive product lines in the industry: a new generation every year, each on or near TSMC's leading node, each pushing performance-per-watt limits competitors needed two or three years to match.
M1 and what it did to expectations
By 2018, Apple had been designing its own iPhone chips for a decade and had built a chip-design organization comparable in size to AMD's. The next move was visible to anyone watching: Apple was going to migrate the Mac from Intel to its own silicon, repeating the 2005 transition.
The M1 launched in November 2020 in the MacBook Air, MacBook Pro 13, and Mac mini. The performance-per-watt was a generational step beyond Intel's mobile chips. A passively-cooled MacBook Air outperformed the actively-cooled Intel MacBook Pro it replaced. Battery life roughly doubled. The chip used the same architectural family (and the same TSMC 5nm process) as the A-series chips, but with more cores and more memory bandwidth — the iPhone team had built a desktop chip by scaling up rather than starting over.
The wider industry reaction was a recognition that the move was now structurally available to anyone with enough scale. Apple had spent twelve years and somewhere in the range of $5-10 billion of cumulative R&D to build the silicon team that delivered M1. That number is large in absolute terms and small for a $400-billion-revenue company that wanted to differentiate at the silicon layer.
The hyperscaler ripple
What Apple proved was that a company with enough volume and enough patience could replace a critical commodity supplier with a custom chip and capture the resulting margin and differentiation. The hyperscalers — Amazon, Microsoft, Google, Meta — were already running enough server volume to make the math work. They started doing exactly this for AI workloads.
Google was earliest. The first TPU shipped internally in 2015, the public Cloud TPU launched in 2017, and by 2024 the TPU v5e and v5p were the chips Google used to train and serve its own models. The TPU is co-designed with TensorFlow and JAX and the Google datacenter network; it is not for sale outside Google Cloud. Amazon followed with Graviton (general-purpose Arm CPUs) in 2018 and Trainium (AI training) and Inferentia (inference) in 2019. Microsoft launched Cobalt (Arm CPU) and Maia (AI accelerator) in 2023. Meta shipped MTIA v1 in 2023 and v2 in 2024.
Each of these chips exists for the same Apple-derived reason: the company has enough internal workload volume that the chip's cost amortizes; the company has enough control of the surrounding software stack to take advantage of architectural choices that a commodity chip cannot make; and the chip-on-paper performance per dollar at internal scale beats the same workload served from NVIDIA. Whether the comparison holds against NVIDIA's actual chips on actual mixed workloads is a separate, more contested question — and the reason NVIDIA still has 80%+ of the AI training market despite four hyperscalers building parallel silicon.
Source: Google "An in-depth look at Google's first Tensor Processing Unit" (2017); AWS reInvent 2018 (Graviton launch); Microsoft Azure blog 2023 (Cobalt/Maia); Meta engineering blog 2024 (MTIA v2).
Strategic read
The Apple Silicon pattern has three preconditions. First, internal workload volume large enough to amortize the chip-design team and the foundry contract — somewhere in the range of one million units per year for a leading-edge node, more for trailing nodes. Second, control of the surrounding software stack tight enough that architectural choices in the chip translate into product-level performance gains. Third, organizational patience to absorb roughly a decade of chip-team build-up before the silicon ships in a flagship product. The third condition is the most often underestimated and the most often missing.
The companies that have completed the pattern at scale are Apple, Google, Amazon, Microsoft, Meta, Tesla (FSD computer), and to a lesser extent NVIDIA (Grace CPU) and AMD (custom semi-custom for Sony, Microsoft consoles). The companies that have started but are not at the same scale yet include OpenAI (rumored to have a custom inference chip in design with Broadcom), Anthropic (using Trainium under their AWS partnership, no own-silicon program disclosed), and a long tail of well-funded AI startups that may or may not get there.
For an investor or operator, the implication is that the customer base for NVIDIA's high-end AI chips is structurally narrowing over time. The hyperscalers are migrating their highest-volume internal workloads to custom silicon, the way Apple migrated off Intel. NVIDIA will retain the workloads that need flexibility (research, frontier training, anything where the model architecture is still moving) and the workloads where the customer cannot justify a $5-10B chip program. That is still a very large business — but it is no longer the same business as 'all AI compute,' and the divergence is one of the most important multi-year trends in the stack.