Skip to content
ContentLora

    Tip: press / anywhere to search.

    Analysis

    Will custom AI chips erode NVIDIA's lead? The 2026 debate

    Custom AI chips grew quickly in 2026. Broadcom's AI semiconductor revenue rose 221% in its quarter to August, and Google began selling TPU systems for customers' own data centers.[1][2] NVIDIA still sells far more: its quarter to July brought $96.2 billion in revenue, and it says its GPUs are "a generation ahead" of ASICs.[3][4] The largest buyers are spreading orders across both kinds of chip.[5][6]

    Editor reviewedUpdated Analysis35 claims · 66% primaryNext-gen computing hardware, Semiconductors, Artificial intelligence, Big Tech
    Show:

    This page weighs whether custom AI chips (ASICs) built by or for cloud and AI companies are eroding NVIDIA’s position in AI computing. It covers developments up to November 7, 2026 and attributes each position to the company or outlet that holds it. It is not investment advice. For the basics of each chip type, read the AI chips explainer first.

    Where things stand in late 2026

    NVIDIA reported revenue of $96.2 billion for the quarter ended July 26, 2026, up 106% from a year earlier, with $89.0 billion from its Data Center segment.[3][7] Broadcom, which develops custom AI accelerators for Google and OpenAI, reported AI semiconductor revenue of $16.7 billion in its quarter ended August 2, 2026, up 221%, and guided $21.7 billion for the next quarter.[8][9][1] Its AI revenue had been $8.4 billion two quarters earlier.[10]

    Alphabet said in October 2026 that Google Cloud revenue rose 103% to $30.7 billion in the third quarter, with TPU system sales a meaningful contributor.[11] It has begun selling TPU systems to a limited number of customers for their own data centers, with most of that revenue due in 2027.[2] Some TPU system contracts require Alphabet to repurchase hardware under specified conditions, with up to $6.6 billion in potential payments.[12]

    The case that custom chips are gaining

    Big buyers have made large, multi-year commitments to non-NVIDIA designs. Broadcom has a long-term agreement to develop and supply Google’s future TPUs through up to 2031, and said Anthropic will access about 3.5 gigawatts of TPU capacity through Broadcom from 2027.[8][13] OpenAI and Broadcom plan to deploy 10 gigawatts of OpenAI-designed accelerators between the second half of 2026 and the end of 2029.[9] Meta says it already runs hundreds of thousands of its own MTIA chips for inference and plans four new generations within two years.[14][15] AWS made its 3nm Trainium3 generally available in December 2025.[16]

    Google’s eighth-generation TPU, announced in April 2026, split into a training chip and an inference chip; Google claims up to 80% better inference performance per dollar than its previous generation.[17][18] Meta reportedly agreed in February 2026 to rent Google Cloud TPUs.[19] TrendForce projected that ASIC-based servers would reach 27.8% of AI server shipments in 2026, their highest share since 2023, growing faster than GPU-based systems.[20]

    On this view, the economics have shifted. Inference is a growing share of AI work, and companies running their own models at huge scale can tune a chip to those models. Each gigawatt committed to a custom design is capacity NVIDIA does not supply. Google selling TPU systems outright removes a barrier, since buyers no longer have to rent them inside Google Cloud.

    The case that NVIDIA’s lead holds

    NVIDIA responded to the November 2025 reports of Meta’s TPU talks by saying it is “a generation ahead of the industry,” the only platform that runs every AI model, and that it offers more performance, versatility and fungibility than ASICs designed for specific frameworks or functions.[21][4] TrendForce’s 2026 projection still had GPUs at about 69.7% of AI server shipments.[20] Tom’s Hardware estimated that a single TPU 8t or 8i chip has roughly a third of the raw compute of NVIDIA’s Vera Rubin or AMD‘s MI455X.[22]

    Switching is costly, according to trade press. The Register noted that TPUs use optical circuit switching rather than packet switches and do not run PyTorch natively.[23] Tom’s Hardware wrote that NVIDIA’s CUDA platform is so widely used that developers find it harder to move to Google’s alternatives.[24] Google’s own guidance says GPUs remain the better fit for models with many custom operations.[25]

    On this view, custom chips mostly serve the few companies that design them, for the workloads they were tuned for. NVIDIA’s revenue is still several times Broadcom’s AI revenue, and GPUs keep their value because they can be resold or reused for any model.

    The case that buyers will run both

    Google told The Register it sees accelerating demand for both TPUs and NVIDIA GPUs and is committed to supporting both.[6] At Cloud Next 2026 it announced NVIDIA Vera Rubin instances alongside TPU 8.[26] Meta describes a portfolio approach that sources silicon from several vendors while keeping MTIA at the centre.[5] In February 2026 it announced deals to buy NVIDIA’s Vera Rubin GPUs and AMD’s MI400 chips, and reportedly agreed to rent TPUs as well.[19][27] Anthropic says it uses Google TPUs, Amazon Trainium and NVIDIA GPUs, and has also agreed to deploy AMD GPUs.[28][29] OpenAI has a separate 6-gigawatt agreement with AMD alongside its Broadcom chip.[30]

    On this view, the question is not replacement but bargaining power. Buyers that can credibly shift work to their own chips can press GPU suppliers on price and allocation, even while total GPU purchases keep rising. Both GPU and custom-chip makers also compete for the same TSMC capacity and HBM memory, covered in why AI chip supply is concentrated.

    What to watch

    These are forecasts by others and dated signposts, not predictions of ours:

    • Broadcom guided AI semiconductor revenue of $21.7 billion for its fiscal fourth quarter.[1]
    • NVIDIA guided revenue of $108.0 billion, plus or minus 2%, for its fiscal third quarter, assuming no Data Center compute revenue from China.[31]
    • Google said TPU 8t and 8i would be generally available later in 2026.[32]
    • Broadcom and OpenAI targeted first deployments of OpenAI-designed accelerators in the second half of 2026.[9]
    • TrendForce projected that Google’s TPU shipments would stay the highest among cloud providers, growing more than 40% in 2026.[33] The Motley Fool, citing TrendForce, reported a 44% rise for Broadcom-made custom processors against 16% for GPUs.[34]
    • Meta plans to bring MTIA 400, 450 and 500 into service for generative AI inference into 2027.[15]
    • Anthropic’s 3.5 gigawatts of TPU capacity through Broadcom starts in 2027 and depends on its commercial success.[13]

    Competing views

    Custom chips are taking real share

    Broadcom's AI semiconductor revenue more than tripled year on year in its latest quarter, Google now sells TPU systems outright, and Anthropic, OpenAI and Meta have committed gigawatts or hundreds of thousands of chips to non-NVIDIA designs. TrendForce expects ASIC servers to reach their highest share since 2023.[1][2][13][9][14][20]

    NVIDIA's lead holds

    NVIDIA says it is a generation ahead and that GPUs are more versatile and fungible than ASICs. GPUs are still about 70% of AI server shipments, single TPU chips trail its GPUs on raw compute, and CUDA and PyTorch habits make switching costly.[4][3][20][22][24][23]

    Buyers will run both

    Google, Meta and Anthropic all say they use several chip platforms. Meta announced deals to buy NVIDIA and AMD GPUs in the same month it reportedly agreed to rent TPUs, and Google offers NVIDIA's Vera Rubin next to its own TPUs.[6][5][27][28][26]

    Next in the crash course

    More in Computing hardware

    Developing in AI & Computing

    Go deeper

    Questions readers ask

    Are custom AI chips replacing NVIDIA GPUs?

    Not in volume so far. TrendForce projected GPUs would be about 69.7% of AI server shipments in 2026 and ASIC-based servers 27.8%, though ASIC shipments are growing faster.[20]

    Which companies make custom AI chips?

    Google designs TPUs with Broadcom, Amazon makes Trainium, Meta makes MTIA chips with deployment in the hundreds of thousands, and OpenAI is designing accelerators that Broadcom will help develop and deploy from the second half of 2026.[8][16][14][9]

    What did NVIDIA say about Google's TPUs?

    After reports in November 2025 that Meta was in TPU talks, NVIDIA said it was delighted by Google's success and still supplies Google, but that it is a generation ahead of the industry and offers more performance, versatility and fungibility than ASICs.[21][4]

    Did Meta sign a deal for Google TPUs?

    SiliconANGLE, citing The Information, reported in February 2026 that Meta agreed a multibillion-dollar deal to rent Google Cloud TPUs, while a separate purchase deal had not been reached.[19]

    Why do companies build their own AI chips?

    Meta says its MTIA chips are more cost-efficient than general-purpose chips for its own workloads. OpenAI's Greg Brockman said building its own chip lets OpenAI embed what it has learned from frontier models into the hardware.[14][35]

    Sources

    · 35 claims from 18 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      Broadcom reported AI semiconductor revenue of $16.7 billion for its fiscal third quarter of 2026 (ended August 2, 2026), up 221% year over year, and guided $21.7 billion for the fourth quarter. confirmedas of 2026-09-02

    2. [2]

      Alphabet says Google Cloud generates product revenue primarily from selling TPU systems, and that it has signed a limited number of agreements to supply TPU systems for customers' on-premises infrastructure, recognizing revenue from them since the second quarter of 2026 with most to be recognized in 2027. confirmedas of 2026-09-30

    3. [3]

      NVIDIA reported revenue of $96.2 billion for the second quarter of fiscal 2027 (ended July 26, 2026), up 106% from a year earlier, with GAAP and non-GAAP gross margins both at 75.0%. confirmedas of 2026-07-26

    4. [4]

      On November 25, 2025, Nvidia posted on X that it was "delighted by Google's success" and continues to supply Google, but that it is "a generation ahead of the industry" and offers greater performance, versatility and fungibility than ASICs designed for specific AI frameworks or functions. confirmedas of 2025-11-25

    5. [5]

      Meta describes a portfolio approach, sourcing silicon from a range of industry leaders while keeping MTIA at the centre of its AI infrastructure, and says it can release a new chip every six months or less. confirmedas of 2026-03-11

    6. [6]

      Google told The Register it was seeing accelerating demand for both its custom TPUs and Nvidia GPUs and was committed to supporting both. reportedas of 2025-11-25

    7. [7]

      In the second quarter of fiscal 2027, NVIDIA's Data Center revenue was $89.0 billion, up 117% year over year, and Edge Computing revenue was $7.2 billion, up 27% year over year. confirmedas of 2026-07-26

    8. [8]

      In April 2026 Broadcom disclosed a long-term agreement to develop and supply custom TPUs for Google's future TPU generations, plus a supply agreement for networking and other components for Google's next-generation AI racks, through up to 2031. confirmedas of 2026-04-06

    9. [9]

      In October 2025 OpenAI and Broadcom announced a collaboration for 10 gigawatts of OpenAI-designed AI accelerators, with Broadcom deploying racks of accelerators and Ethernet networking from the second half of 2026 through the end of 2029. confirmedas of 2025-10-13

    10. [10]

      Broadcom reported AI revenue of $8.4 billion for its fiscal first quarter of 2026, up 106% year over year, driven by demand for custom AI accelerators and AI networking. confirmedas of 2026-03-04

    11. [11]

      Alphabet reported that Google Cloud revenue grew 103% to $30.7 billion in the third quarter of 2026, with TPU systems sales a meaningful contributor, and Sundar Pichai said Google was seeing strong demand for TPU systems across a wide range of customers. confirmedas of 2026-10-28

    12. [12]

      Alphabet says some TPU system contracts require it to repurchase hardware under specified conditions, with maximum potential future payments of $6.6 billion as of September 30, 2026. confirmedas of 2026-09-30

    13. [13]

      Broadcom said Anthropic will access about 3.5 gigawatts of next-generation TPU-based compute through Broadcom beginning in 2027, with consumption dependent on Anthropic's continued commercial success. confirmedas of 2026-04-06

    14. [14]

      Meta says it deploys hundreds of thousands of MTIA chips for inference across content and ads, and that MTIA is more cost-efficient than general-purpose chips for those workloads. confirmedas of 2026-03-11

    15. [15]

      In March 2026 Meta said it is developing four new MTIA chip generations, the MTIA 300, 400, 450 and 500, within two years; MTIA 300 is in production for ranking and recommendation training, and the later chips will mainly serve generative AI inference into 2027. confirmedas of 2026-03-11

    16. [16]

      In December 2025 AWS made EC2 Trn3 UltraServers generally available, powered by Trainium3, its first 3nm AI chip, scaling to 144 chips and up to 362 FP8 petaflops, with Anthropic among the customers using Trainium. confirmedas of 2025-12-02

    17. [17]

      On April 22, 2026, at Google Cloud Next, Google introduced its eighth-generation TPU as two separate chips, TPU 8t for training and TPU 8i for inference. confirmedas of 2026-04-22

    18. [18]

      Google claims TPU 8t delivers up to 2.7 times the training performance per dollar of Ironwood and TPU 8i up to 80% better inference performance per dollar, and that both offer up to twice Ironwood's performance per watt. confirmedas of 2026-04-22

    19. [19]

      In February 2026 The Information reported, according to SiliconANGLE, that Meta had agreed a multibillion-dollar deal to rent Google Cloud TPUs to train and run its models, while a separate deal to buy TPUs for its own data centers had not been reached. reportedas of 2026-02-26

    20. [20]

      TrendForce projected in January 2026 that GPUs would account for about 69.7% of AI server shipments in 2026 and ASIC-based servers 27.8%, the highest ASIC share since 2023, with ASIC server shipments growing faster than GPU-based systems. reportedas of 2026-01-22· forecast

    21. [21]

      In November 2025 The Information reported that Meta was in talks to use Google's TPUs in its own data centers from 2027, which The Register and Tom's Hardware relayed. reportedas of 2025-11-25

    22. [22]

      Tom's Hardware, citing Google's technical deep dive, reported TPU 8t at 12.6 FP4 petaflops with 216 GB of HBM3e and TPU 8i at 10.1 FP4 petaflops with 288 GB of HBM3e, roughly a third of the per-chip compute of Nvidia's Vera Rubin or AMD's MI455X. reportedas of 2026-04-27

    23. [23]

      The Register argued that TPU clusters use optical circuit switching rather than packet switches and do not run PyTorch natively, requiring the PyTorch/XLA translation layer, which makes a switch from GPUs harder for a PyTorch user such as Meta. reportedas of 2025-11-25· interpretation

    24. [24]

      Tom's Hardware wrote that Nvidia's CUDA platform is widely used across the industry, making it harder for developers to switch to Google's alternatives. reportedas of 2025-11-25· interpretation

    25. [25]

      Google's guidance says TPUs suit models dominated by matrix computations, large models that train for weeks or months and models with very large embeddings, while GPUs suit models with many custom PyTorch or JAX operations, and TPUs are not suited to high-precision arithmetic or code with frequent branching. confirmedas of 2026-10-26

    26. [26]

      Google also announced Nvidia Vera Rubin NVL72 instances on the same Virgo Network fabric at Cloud Next 2026. reportedas of 2026-04-27

    27. [27]

      SiliconANGLE reported that in February 2026 Meta also announced deals to buy Nvidia's Vera Rubin GPUs and billions of dollars of AMD AI chips, including the Instinct MI400 series. reportedas of 2026-02-26

    28. [28]

      In October 2025 Anthropic said it would expand to up to one million Google TPUs and that it uses three chip platforms: Google TPUs, Amazon Trainium and NVIDIA GPUs. confirmedas of 2025-10-23

    29. [29]

      In July 2026 Anthropic agreed to deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs in Helios rack-scale systems, with the first gigawatt beginning in the first half of 2027. confirmedas of 2026-07-22

    30. [30]

      In October 2025 AMD and OpenAI agreed to deploy 6 gigawatts of AMD Instinct GPUs across multiple generations, starting with 1 gigawatt of MI450 GPUs in the second half of 2026; AMD issued OpenAI a warrant for up to 160 million AMD shares that vests with deployment milestones and AMD share-price targets. confirmedas of 2025-10-06

    31. [31]

      NVIDIA guided third-quarter fiscal 2027 revenue of $108.0 billion, plus or minus 2%, and assumes no Data Center compute revenue from China in that outlook. confirmedas of 2026-08-26· forecast

    32. [32]

      Google said in April 2026 that both eighth-generation TPUs would be generally available later in 2026. confirmedas of 2026-04-22· forecast

    33. [33]

      TrendForce said Google was partnering with Broadcom on the TPU v7p (Ironwood) and projected that Google's TPU shipments would stay the highest among cloud providers, growing more than 40% in 2026. reportedas of 2026-11-07· forecast

    34. [34]

      The Motley Fool reported that TrendForce expected shipments of the custom AI processors Broadcom makes to grow 44% in 2026, against 16% growth in GPU shipments. reportedas of 2025-12-31· forecast

    35. [35]

      OpenAI president Greg Brockman said that by building its own chip OpenAI can embed what it has learned from frontier models directly into the hardware. confirmedas of 2025-10-13

    Revision history (2)
    1. Re-checked; no new NVIDIA, Broadcom or Google results since the page was created. NVIDIA reports on 17 November.
    2. Page created, covering developments up to November 7, 2026.

    Created Nov 7, 2026. Last reviewed by an editor on Nov 8, 2026. Next scheduled review: Nov 21, 2026.

    Cite this page

    "Will custom AI chips erode NVIDIA's lead? The 2026 debate." ContentLora, updated Nov 14, 2026. https://contentlora.com/analysis/custom-ai-chips-vs-gpus-debate

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.