Skip to content
FAQ
AI & ML Analysis

AI Week, Mar 16–22, 2026: A $1 trillion order forecast, a terawatt chip fab and cheaper tokens

Nvidia put its Blackwell and Vera Rubin orders at $1 trillion through 2027, OpenAI shipped GPT-5.4 mini and nano, Accenture and Microsoft launched an AI engineering practice, Uber plans to invest up to $1.25 billion in Rivian for robotaxis, and Elon Musk unveiled TERAFAB in Austin.

6 min read

Written on November 15, 2026, covering news from Mar 16–22, 2026.

One week in March put several answers to the AI compute squeeze on stage. Nvidia projected $1 trillion in purchase orders through 2027, OpenAI released smaller and cheaper models, and Elon Musk said outside chipmakers could not grow fast enough for his companies, so he would build his own plant.

Monday, March 16

At GTC in San Jose, Nvidia CEO Jensen Huang said he expects purchase orders across Blackwell and Vera Rubin to hit $1 trillion through 2027. Techloy quoted him describing “through 2027 at least $1 trillion” in high-confidence demand and purchase orders for Blackwell, Rubin and future platforms, and said this doubles the $500 billion projection he gave at GTC 2025. CNBC noted that after Nvidia’s earnings report the month before, finance chief Colette Kress said growth this year should top what that earlier estimate included.

Huang framed the demand in terms of customer capacity: “If they could just get more capacity, they could generate more tokens, their revenues would go up”. Speaking to supply chain partners in the room, he said, “I know why you’re not impressed because all of you had record years,” and added that he is “certain computing demand will be much higher than that”. Nvidia shares rose about 2% that day.

For context, CNBC put Nvidia’s value at about $4.5 trillion, the largest of any public company, and said the company guided in February to roughly $78 billion in revenue this quarter, up about 77% year over year, after 11 straight quarters of growth above 55%.

The hardware news:

  • Vera Rubin is scheduled to roll out later this year; Nvidia claims the system, built from 1.3 million components, delivers 10 times the performance per watt of Grace Blackwell.
  • Huang unveiled the Nvidia Groq 3 Language Processing Unit, Nvidia’s first chip from the startup; Nvidia bought most of Groq in December via a $20 billion asset deal, the biggest it has ever done, and the chip is expected to ship in the third quarter.
  • The Groq 3 LPX rack holds 256 LPUs and is designed to sit next to Vera Rubin racks; Huang said it can raise the tokens per watt of Rubin GPUs by 35 times.
  • Huang showed a prototype of Kyber, a rack architecture with 144 GPUs in vertically mounted compute trays, which will come in Vera Rubin Ultra, expected to ship in 2027.

On the Groq chips, Huang said: “We united, unified two processors of extreme differences, one for high throughput, one for low latency”.

On software, Huang introduced NemoClaw, a reference stack for OpenClaw, the agent project launched in January by Austrian developer Peter Steinberger, aimed at making it “enterprise ready”. CNBC noted Steinberger joined OpenAI last month.

In automotive, Huang said Uber will roll out a fleet running Nvidia’s Drive AV software in 28 cities on 4 continents by 2028, with Los Angeles and San Francisco going first next year. He also said level 4 autonomous vehicles are being developed on Drive Hyperion by Nissan, BYD, Geely, Isuzu and Hyundai, while Isuzu and Japan’s Tier IV are building autonomous buses on the platform with Nvidia’s AGX Thor chip.

Tuesday, March 17

OpenAI released GPT-5.4 mini and GPT-5.4 nano, which it called its “most capable small models yet”. Both arrived two weeks after GPT-5.4 itself.

OpenAI says mini improves on GPT-5 mini in coding, reasoning, multimodal understanding and tool use, runs more than 2x faster, and approaches the full GPT-5.4 on evaluations including SWE-Bench Pro and OSWorld-Verified. Simon Willison noted that OpenAI’s self-reported benchmarks show nano beating the earlier GPT-5 mini when run at maximum reasoning effort. OpenAI recommends nano for classification, data extraction, ranking and coding subagents doing simpler supporting work.

Availability differs by model. Mini is live in ChatGPT, Codex and the API; Free and Go users reach it through the “Thinking” option in the + menu, while other users get it as a rate limit fallback for GPT-5.4 Thinking. Nano is only in the API.

Willison published per-million-token pricing:

  • gpt-5.4: $2.50 input, $15.00 output.
  • gpt-5.4-mini: $0.75 input, $4.50 output.
  • gpt-5.4-nano: $0.20 input, $1.25 output, which he noted undercuts Google’s Gemini 3.1 Flash-Lite at $0.25 and $1.50.

In his test, nano described one photo using 2,751 input tokens and 112 output tokens for 0.069 cents, which he calculated would put his 76,000-photo library at about $52.44. He shipped llm 0.29 with support for the new models.

Wednesday, March 18

Accenture and Microsoft launched a forward deployed engineering (FDE) practice meant to help companies design, build and run AI across their operations. The firms said it will put thousands of AI-skilled engineers directly with clients, with Microsoft supplying the platform and Accenture handling change management, process redesign and global rollout. Joint teams will work with clients using Microsoft’s Frontier Suite, with a stated aim of moving AI from idea to production “in days, not months”.

“AI value does not come from technology access but from the ability to convert it into sustained business impact,” said Manish Sharma, Accenture’s chief strategy and services officer. Microsoft chief business officer Nick Parker said, “What customers are asking us for is a clear path to business outcomes powered by AI”.

Saturday, March 21

On Saturday night at the Seaholm Power Plant in Austin, Elon Musk unveiled TERAFAB, a joint chip effort between Tesla, SpaceX and xAI, calling it “the most epic chip-building exercise in history by far”. Greg Abbott, the governor of Texas, sat in the audience.

The price tag differs by outlet: KXAN, carrying an Austin Business Journal report, put it at $20 billion, while Teslarati described a $20-25 billion joint venture. Musk’s goal is more than a terawatt of AI compute per year, which KXAN said equals about 100 million to 200 million AI chips. Musk said current AI compute output is “roughly 20 gigawatts per year” and that the rest of Earth’s output is about 2% of what his companies need.

According to Teslarati, the plan covers design, lithography, fabrication, memory, advanced packaging and testing in one place, and at full capacity would reach roughly 70% of TSMC’s global output. Teslarati also reported that Musk said 80% of the compute would go to orbital AI satellites, and described the terawatt goal as divided between Tesla’s AI5 inference chips, for vehicles and Optimus, and D3 chips built for orbital satellites. Musk credited Samsung, TSMC and Micron but said “there’s a maximum rate at which they’re comfortable expanding”.

Where it will be built is less settled. KXAN reported Musk saying the plant would go on the Tesla campus in eastern Travis County. But in a Sunday post on X, Musk said the Austin-area facility is one part of the project and will focus on chip design, while the main facility would need thousands of acres, with multiple locations under consideration. KUT said Musk gave no further details on the Austin footprint or on construction and production timelines.

Also this week

Reported on March 19: Uber plans to put up to $1.25 billion into Rivian under a deal targeting as many as 50,000 robotaxis across several countries through 2031. Uber or its fleet partners are expected to buy 10,000 autonomous R2 vehicles, with an option for up to 40,000 more starting in 2030. An initial $300 million, equal to about 19.55 million shares, is expected soon after signing, subject to regulatory approval, with 4 further tranches tied to milestones. The robotaxis are expected in 25 cities across the U.S., Canada and Europe, beginning with San Francisco and Miami in 2028. Rivian closed 3% higher that day while Uber fell 1%. CNBC noted that many companies, Uber included, have previously missed robotaxi targets.

What to Watch

  • Kyber and Vera Rubin Ultra: Nvidia expects that system to ship in 2027, which will test whether the $1 trillion order view holds up.
  • TERAFAB’s actual site and schedule: Musk has not given a construction or production timeline, and the reports conflict on whether the Austin campus hosts the fab itself or a design operation.
  • Uber’s two robotaxi tracks: Nvidia-powered fleets starting in Los Angeles and San Francisco next year, and Rivian R2 launches in San Francisco and Miami in 2028, with Rivian’s later tranches depending on milestones.
  • My read is that pricing will drive nano’s adoption more than benchmarks; the gap Willison flagged against Gemini 3.1 Flash-Lite is the number to track.

This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.

Nvidia OpenAI GPT-5.4 TERAFAB Tesla Uber Rivian Accenture
Share