Model makers make chips, and vice versa

The contest over artificial intelligence extends beyond the model answering on screen. On 24 June 2026, OpenAI and Broadcom unveiled Jalapeño, the ChatGPT provider's first custom accelerator. Its samples were already running machine learning workloads in the lab.[1]

In October, NVIDIA expanded its model range with Nemotron 3.5 Bolt, designed to run on local machines.[2] The machine supplier also provides intelligence to run on them. OpenAI is moving in the opposite direction: it wants to design the engine that will run its models.

Both moves seek greater control over the business. To understand why, it helps to look at what is happening to the price of the capability they sell.

Intelligence is getting cheaper and can be reproduced

A capability that was exceptional yesterday can soon become a cheap service. Stanford documented a more than 280-fold fall in the cost of inference needed to reach GPT-3.5-level performance between November 2022 and October 2024.[3] Inference means running an already trained model to obtain answers.

DeepSeek R1 made that pressure visible. Its launch price was $0.55 per million input tokens and $2.19 per million output tokens, compared with the $15 and $60 listed for o1. The price per token was roughly one twenty-seventh of o1's.[4][5] Tokens are the fragments of text that a model processes and generates.

Capabilities can also be transferred. DeepSeek trained smaller versions using examples produced by R1: distillation means teaching a model through another model's answers.[6] Publishing the weights—the numerical values learned during training—also allows people to download and run them under the terms of their licence.[2] That combination puts pressure on model vendors to seek advantages elsewhere in the chain.

Power seeks factories, memory and distribution

Making chips and reaching users require resources that are harder to reproduce than a weights file. NVIDIA reported a gross margin of 75% in its second quarter of fiscal 2027, which ended in July 2026.[7] That position in turn depends on suppliers: TSMC said in October that the demand it assessed still exceeded supply, while SK hynix announced an expansion of HBM4.[8][9] HBM is high-bandwidth memory: it feeds data to the processor so it can work.

Large buyers design their own chips to tailor the machine to the workload and negotiate supply. Google develops TPUs, its tensor processing units, chips specialised for AI. Amazon develops Trainium; Microsoft says Maia 200 is already serving production workloads.[10][11][12]

NVIDIA also seeks access to users. In September it announced an agreement to acquire Hugging Face, the model distribution platform.[13] As the giants try to control those pieces, another trend is expanding the options for people using AI.

More intelligence fits into less space

Local capability grows in two ways: models that do more at a smaller size, and processors that run them better. Xiao and colleagues estimated that maximum capability density doubled every 3.5 months in their sample of 51 open base models.[14] Density expresses how much larger a reference model would need to be to match the performance of the model being studied.

Alibaba offers an example: it reports 79.1 points on the MMLU-Pro knowledge test for Qwen3.5-4B, compared with 71.1 for Qwen2.5-72B, from roughly eighteen months earlier.[15][16] The first model's language component has about four billion parameters; the second has 72.7 billion. Parameters are the numerical quantities a model learns.

Epoch AI estimated a gap of roughly 6–12 months, depending on the test, between the most capable models of the day and models that fit entirely on an RTX 4090 or 5090 consumer graphics card. Its calculation assumes four-bit quantization and an 8K-token context, as well as auxiliary memory.[17] Quantization means storing the model's numbers using fewer bits; context is the text it handles at once. NVIDIA combines that technique with compression and additional training in Bolt.[18]

Hardware is advancing too. Google credits TPU 8i with up to twice Ironwood's performance per watt. Apple increases maximum memory bandwidth from 546 gigabytes per second on M4 Max to 614 on M5 Max: more data reaches the processor.[19][20][21] Both curves expand the work that can move out of the data centre.

AI comes to the law firm, with costs to calculate

That convergence makes it possible to consider running in-house tasks on an office machine. NVIDIA announced a DGX Spark with 64 gigabytes of memory, intended for models with up to 100 billion parameters; Apple announced Mac Studio M5 Ultra with up to 512 gigabytes.[22][23]

The starting point is narrowly defined work. An academic study found that Mistral 7B outperformed Llama 70B on two legal classification tests and lost on others.[24] Classifying documents, extracting fields and searching in-house files are candidates for evaluation. The most advanced frontier remains above, in the large data centres; Artificial Analysis evaluations still show a gap between it and the best models with published weights.[25]

The economics must be calculated for each workflow. In this example, a cheap API wins. An API is the connection between an application and the provider's service. A $5,000 machine depreciated over three years, including energy and operation, costs about $268 a month. Ten thousand medium-sized tasks a month cost about $8 at Gemini 2.5 Flash-Lite's July 2025 price.[26][27] This is an economic snapshot using a historical price: more compact models, more efficient chips and external services that also keep getting cheaper mean the comparison must be revisited.

Start with one workflow and decide through testing

A useful decision starts with a specific job: what goes in, what output is acceptable and who reviews it. Quality, time and cost per accepted task must be measured. Those tests determine what to run in-house, what to buy through an API and what to combine.

Control also requires tracking the data. The General Data Protection Regulation, GDPR, applies to documents, logs and tools throughout processing.[28][29] A machine in the office can send information outside if it uses external services.

The power a firm can regain starts with that choice: deciding which intelligence it uses, where it runs it and whom it entrusts with its data.

Sources

Sources from the research, accessed on 06-11-2026. Document dates and manufacturer attributions are retained; entries without their own date may change.

  1. OpenAI / Broadcom, 24-06-2026, primary announcement, OpenAI and Broadcom unveil LLM-optimized inference chip. https://openai.com/index/openai-broadcom-jalapeno-inference-chip/
  2. NVIDIA, 15-10-2026 according to ReleaseDate, mutable model/licence card, NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB. https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB
  3. Stanford HAI, April 2025, primary report, AI Index Report 2025. https://hai.stanford.edu/ai-index/2025-ai-index-report
  4. DeepSeek, 20-01-2025, primary announcement and historical pricing, DeepSeek-R1 Release. https://api-docs.deepseek.com/news/news250120/
  5. OpenAI, undated; identifier o1-2024-12-17, mutable model card accessed at the research cutoff, o1; listed as retired. https://developers.openai.com/api/docs/models/o1
  6. DeepSeek, January 2025; mutable repository, primary results and licences, DeepSeek-R1. https://github.com/deepseek-ai/DeepSeek-R1
  7. NVIDIA, 26-08-2026, company results, NVIDIA Announces Financial Results for Second Quarter Fiscal 2027. https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-Second-Quarter-Fiscal-2027/default.aspx
  8. TSMC / LSEG transcript, 15-10-2026, primary earnings call, Q3 2026 Earnings Call. https://investor.tsmc.com/japanese/encrypt/files/encrypt_file/reports/2026-10/eba4fac72873bf546b669bd2b5d973df5c93215d/TSMC%203Q26%20Transcript.pdf
  9. SK hynix, 28-10-2026, company results, SK hynix Announces 3Q26 Financial Results. https://news.skhynix.com/en/q3-2026-business-results/
  10. Google, 22-04-2026, primary technical announcement, Our eighth generation TPUs: two chips for the agentic era. https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/eighth-generation-tpu-agentic-era/
  11. Amazon, Q1 2026 results, 29-04-2026; no separate date identified for the article, CEO statement, Andy Jassy weighs in on the rapid growth of Amazon's chips business. https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-amazon-chips-business-q1-2026-earnings
  12. Microsoft, 12-10-2026, primary announcement, Microsoft to expand Azure Maia system to multi-gigawatt deployments. https://blogs.microsoft.com/blog/2026/10/12/microsoft-to-expand-azure-maia-system-to-multi-gigawatt-deployments/
  13. NVIDIA, 03-09-2026, announced agreement, NVIDIA to Acquire Hugging Face. https://blogs.nvidia.com/blog/nvidia-to-acquire-hugging-face/
  14. Chaojun Xiao and colleagues, 06-11-2025, academic empirical estimate, peer-reviewed study on LLM capability density, Nature Machine Intelligence. https://www.nature.com/articles/s42256-025-01137-0
  15. Alibaba / Qwen, March 2026; mutable card accessed 06-11-2026, manufacturer evaluation, Qwen3.5-4B. https://huggingface.co/Qwen/Qwen3.5-4B
  16. Alibaba / Qwen, 19-12-2024; model released 19-09-2024, technical report and manufacturer evaluation, Qwen2.5 Technical Report, table 6. https://arxiv.org/html/2412.15115v1 — Comparison of published MMLU-Pro results: protocols and reasoning budgets were not matched across generations.
  17. Somala and Emberson / Epoch AI, 15-08-2025; data updated 18-11-2025, independent estimate, Consumer GPU model gap. https://epoch.ai/data-insights/consumer-gpu-model-gap — Assumptions: all weights in graphics card memory, four-bit quantization, an 8K-token context, an eight-bit key-value cache (KV INT8) and around 2 GiB of additional overhead. GiB stands for gibibyte, a binary unit of memory. Limits of 28 billion parameters for RTX 4090 and 40 billion for RTX 5090. Average gaps by test: 7.4, 7.3, 6.3 and 12.4 months. Fitted score trends, not a measurement of usable speed.
  18. NVIDIA Research, 15-10-2026, technical explanation and manufacturer measurements, Nemotron 3.5 Bolt 24GB. https://research.nvidia.com/labs/nemotron/Nemotron-3.5-Bolt-24GB/
  19. Google Cloud, 22-04-2026, manufacturer's technical statement, Inside the eighth-generation TPU: An architecture deep dive. https://cloud.google.com/blog/products/compute/tpu-8t-and-tpu-8i-technical-deep-dive/?hl=en
  20. Apple, 30-10-2024, manufacturer specifications, Apple introduces M4 Pro and M4 Max. https://www.apple.com/newsroom/2024/10/apple-introduces-m4-pro-and-m4-max/
  21. Apple, 03-03-2026, manufacturer specifications, Apple debuts M5 Pro and M5 Max. https://www.apple.com/newsroom/2026/03/apple-debuts-m5-pro-and-m5-max-to-supercharge-the-most-demanding-pro-workflows/
  22. Allen Bourgoyne / NVIDIA Blog, 02-10-2026, announcement, DGX Spark 64GB / local AI. https://blogs.nvidia.com/blog/local-ai-dgx-spark-64gb-sync/
  23. Apple Newsroom, 25-08-2026, announcement, Apple introduces new Mac Studio with M5 Max and M5 Ultra. https://images.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/
  24. Bing Zhang and colleagues, 30-04-2025, peer-reviewed NAACL Industry paper, Evaluating Large Language Models with Enterprise Benchmarks. https://aclanthology.org/2025.naacl-industry.40.pdf
  25. Artificial Analysis, 27-10-2026, methodology and independent evaluation, Artificial Analysis Intelligence Index v5. https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5
  26. Google Developers Blog, 22-07-2025, announcement and historical price, Gemini 2.5 Flash-Lite is now stable and generally available. https://developers.googleblog.com/gemini-25-flash-lite-is-now-stable-and-generally-available/
  27. Own calculation, 06-11-2026, illustrative scenario: 5,000/36 + 0.2 × 8 × 22 × 0.25 + 2 × 60 = 267.69 USD/month; API: 40 × 0.10 + 10 × 0.40 = 8 USD/month. Machine depreciated over 36 months; electricity at 0.25 USD/kWh; 176 hours/month at 0.2 kW; two hours of operation at 60 USD/hour. Workload: 10,000 tasks, 4,000 input tokens and 1,000 output tokens per task. Common integration and human review are excluded for both options; equivalent quality is not assumed. Google's historical July 2025 pricing is documented in the preceding source. Method without an external URL.
  28. European Union, 27-04-2016; OJ EU 04-05-2016, primary legislation, Reglamento (UE) 2016/679. https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=CELEX%3A32016R0679
  29. EDPB, 18-12-2024, primary regulatory guidance, Opinion 28/2024 on certain data protection aspects related to the processing of personal data in the context of AI models. https://www.edpb.europa.eu/system/files/2024-12/edpb_opinion_202428_ai-models_en.pdf

Text by Roberto Ortega, prepared with AI editorial assistance. The Norwegian and English editions are editorial adaptations of the Spanish original.