Shipments of the custom AI chips that Google, Amazon and other cloud giants design for their own data centers will cross 15 million in 2028, surpassing data center GPU shipments, Counterpoint Research estimated in a January 2026 insight. Google’s TPUs will remain the volume backbone, the research firm said.
The forecast resurfaced on 24 October 2026, when Milk Road AI posted a chart, credited to Counterpoint and the Korean broker Hana Securities, showing AI GPU and ASIC shipments rising from roughly 7 million units in 2024 to more than 40 million by 2031.
Business Pill · GPUS
A short explainer of why AI runs on GPUs. It teaches the general idea only and says nothing about custom chips or Counterpointโs forecasts.
The key insight: As we read it, the new element is a date for the crossover: Counterpoint estimates custom AI chips cross 15 million units in 2028 and outship data center GPUs that year, with Googleโs TPUs as the volume backbone even as Googleโs share falls from 64% in 2024 to 52% in 2027.
Counterpoint’s 26 January insight said global AI server compute ASIC shipments for the top 10 players are set to triple between 2024 and 2027. ASICs are chips built for one kind of work, here AI training and inference, rather than general-purpose GPUs.
Research Vice President Neil Shah was quoted: “we estimate that global data center AI Server Compute ASIC shipments will cross the 15-million mark in 2028, surpassing data center GPU shipments.” Shah added that the top 10 AI hyperscalers combined “will deploy more than 40 million AI Server Compute ASIC chips cumulatively during 2024-2028.”
The insight described a market moving from a duopoly, Google with 64% and AWS with 36% of shipments in 2024, to volume ramp-ups from Meta (MTIA) and Microsoft (Maia) by 2027. Google’s share is expected to fall to 52% in 2027, Research Associate David Wu said, while its TPU fleet “will remain the undisputed volume backbone and north star of the industry.”
Counterpoint figure
Value
AI server compute ASIC shipments, top 10 players, 2024 to 2027
about 3x
Data center AI ASIC shipments, 2028
more than 15 million
Cumulative ASICs deployed by the top 10 hyperscalers, 2024 to 2028
more than 40 million
Google share of ASIC shipments, 2024 / 2027
64% / 52%
AWS share of ASIC shipments, 2024
36%
Broadcom share as ASIC design partner, 2027
60%
Marvell design service share, 2027
8%
TSMC share of wafer fabrication for the top 10 players
close to 99%
Source: Counterpoint Research insight, 26 January 2026.
Share of AI server compute ASIC shipments by hyperscaler, top 10 players: Google 64% and AWS 36% in 2024, Google 52% forecast for 2027. Source: Counterpoint Research insight, 26 January 2026.
What Was Posted on X
The AI infrastructure buildout is about to get MUCH bigger than people realize.
According to estimates from Counterpoint Research, AI GPU and ASIC shipments are projected to increase from roughly 7 million units in 2024 to more than 40 million units by 2031. That’s nearly a 6x increase, and what I find particularly interesting is that this growth isn’t expected to come from NVIDIA alone. While $NVDA is expected to remain a major player, companies like $GOOGL, $AMZN, $MSFT, and $META are increasingly developing their own custom AI chips, also known as ASICs. These chips are designed for specific workloads, allowing hyperscalers to optimize performance, reduce costs, and become less dependent on NVIDIA for every piece of compute they deploy.
The rise of custom AI chips doesn’t necessarily mean NVIDIA loses but rather mean the overall AI compute market becomes significantly larger. Hyperscalers can continue purchasing NVIDIA GPUs while simultaneously deploying their own chips for workloads where custom silicon makes more economic sense. And there are implications far beyond the companies designing these processors. Every additional GPU or ASIC needs advanced memory, networking, packaging, and power. As AI systems become more powerful, the amount of HBM required per accelerator can also increase, while connecting tens of thousands of chips creates additional demand for high speed networking and optical interconnects. This is exactly why I remain so bullish on the broader AI infrastructure ecosystem, particularly companies exposed to memory like $MU, networking like $CRDO, and optical connectivity.
Milk Road AI says Counterpoint estimates put AI GPU and ASIC shipments at roughly 7 million units in 2024 and more than 40 million by 2031, with hyperscalers building their own chips alongside Nvidia. The post also names stocks it favours; those views are its own.
Google’s New TPUs Ramp Tenfold
A second Counterpoint insight, of 29 April 2026, projected combined shipments of Google’s TPU v8t and v8e to approach 5 million units in 2028, more than 10x the approximately 400,000 units shipped in 2026. MediaTek, which works with Google on those chips, is positioned to capture 26% of total AI ASIC server compute shipments by 2028, the firm said.
Counterpoint tied the shift to memory costs. Wu said Google’s motivation “lies in HBM procurement economics,” explaining that under the Broadcom turnkey model the ASIC vendor sources the HBM and applies a 15%-20% markup. Associate Director Brady Wang said “achieving the projected volume is contingent upon sufficient advanced packaging capacity, including TSMC CoWoS allocation and Intel EMIB-T yield readiness for Humufish.”
The 2031 Figure
Milk Road AI’s post said that, according to Counterpoint estimates, AI GPU and ASIC shipments are projected to increase from roughly 7 million units in 2024 to more than 40 million units by 2031, “nearly a 6x increase.” The attached chart is labelled in Korean as an AI GPU and ASIC shipment outlook, with the source line Counterpoint and Hana Securities, and splits each year by vendor, from Nvidia, AMD and Intel to Google, Meta, AWS, Microsoft and OpenAI.
Counterpoint lists a Cloud AI CPU, GPU and Compute ASIC Tracker, published 16 August 2026, with forecast shipment data through 2031; the full report requires a subscription. We could not read it, so the 2031 numbers rest on Milk Road AI’s post and the chart it shared. Note the two 40 million figures differ: Counterpoint’s public insight counts ASICs deployed cumulatively from 2024 to 2028, while the chart shows GPUs and ASICs shipped in the single year 2031.
Connecting the Coverage
Our TPUv7 piece reported SemiAnalysis saying on 15 October that Google’s TPUv7 showed 1.67x better performance per dollar than Nvidia’s GB300 NVL72 in a preview benchmark. Our PyTorch piece reported Google open-sourcing TorchTPU on 20 October, a day after Amazon released TorchNeuron for its Trainium chips.
On the GPU side, our H100 rent piece reported Silicon Data’s H100 neo-cloud index at $2.88 per GPU-hour on 24 October, up from $2.00 on 9 December 2025. Our sellout piece reported Micron saying its strategic customer agreements run through 2030. And our substrate piece reported that Tesla and Nvidia will put about $2.8 billion into Samsung Electro-Mechanics chip-substrate lines, per a Korean newspaper report; none of the companies confirmed the amounts.
What is new here: a dated forecast that custom chips overtake GPUs on units in 2028, with Google’s own TPU ramp sized by the same firm.
The Structural Read
As we read it, the growth is spread across more buyers: Counterpoint expects Meta (MTIA) and Microsoft (Maia) to ramp by 2027, after a 2024 market split between Google and AWS.
As we read it, memory is part of the design choice: Counterpoint says Google working with MediaTek avoids the 15%-20% HBM markup in the Broadcom turnkey model.
As we read it, the 2031 chart and Counterpointโs public insights measure different things: the chart shows GPUs and ASICs shipped in one year, the insight counts ASICs deployed cumulatively from 2024 to 2028.
Neil Shah, Counterpoint Research, 26 January 2026
“we estimate that global data center AI Server Compute ASIC shipments will cross the 15-million mark in 2028, surpassing data center GPU shipments.”
Three Implications
FOR CHIP BUYERS Counterpoint expects Meta and Microsoft to ramp their own chips by 2027, alongside Google and AWS.
FOR DESIGN PARTNERS Counterpoint projects Broadcom at 60% of design-partner share in 2027 and MediaTek at 26% of shipments by 2028.
FOR MEMORY AND PACKAGING Counterpoint says the TPU ramp is contingent on TSMC CoWoS allocation and Intel EMIB-T yield readiness.
The framework puts it this way: “Layer 3 — Silicon. Three competitive categories within one layer: generalist substrate (NVIDIA), hyperscaler custom (TPU, Trainium, MTIA), merchant custom challengers (Cerebras, Groq, SambaNova).”
As we read it, Counterpointโs forecast is a unit count from the hyperscaler-custom category of that layer. The framework does not discuss the forecast.
What Is Not Established
These are Counterpoint forecasts, not reported shipments. The insights count units, not dollars or computing power, and one ASIC and one GPU are not equal in capability. Counterpoint’s public insights give no 2027 AWS share and no GPU shipment total. The 2031 figures come from a broker chart shared on X; we did not see the underlying report. We did not contact Counterpoint, Hana Securities or Milk Road AI.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Counterpoint estimates that custom AI chips will cross 15 million units in 2028 and outship data center GPUs, with Google’s TPUs as the volume leader even as its share falls from 64% in 2024 to 52% in 2027.
A note on sourcing. This piece rests on Counterpoint Research’s public insights of 26 January and 29 April 2026, its 16 August 2026 tracker listing, Milk Road AI’s X post and chart of 24 October 2026, and our own earlier pieces, all read on 24 October 2026. Counterpoint’s subscription tracker was not read. We did not contact Counterpoint, Hana Securities or Milk Road AI.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
