March had a major model launch, a new chip platform, and the largest private funding round ever. The story I keep coming back to is a shutdown.
GPT-5.4 puts computer use in the base model
OpenAI released GPT-5.4 on March 5. The headline feature is native computer use: the model reads screenshots, clicks, types, and moves between applications without a separate agent wrapper. It also absorbs the coding strengths of GPT-5.3-Codex, so OpenAI now has one flagship for chat, code, and operating software instead of three. GPT-5.4 mini and nano followed on March 17, with mini going to free users.
The consolidation matters more than any benchmark for teams building on the API. Fewer models means fewer routing decisions. It also tells you OpenAI expects most paid usage to be agentic soon.
Nvidia designs for inference
At GTC, Nvidia announced that the Vera Rubin platform is in full production: seven chips including the Vera CPU, Rubin GPU, NVLink 6, BlueField-4, Spectrum-6, and a Groq 3 LPU for low-latency inference. Partner systems ship in the second half of the year.
The Groq LPU is the detail to notice. Agents make many small sequential calls, and that workload needs a different latency profile than training. Nvidia is now shaping its roadmap around inference.
OpenAI kills Sora
On March 24, OpenAI told developers the Sora app would close on April 26 and the API on September 24. It gave no reason. Reporting pointed to compute cost and a shift toward enterprise and coding products.
This is the most useful signal of the month. Sora was a flagship launch and one of the most recognizable AI products in the world. OpenAI cut it anyway once the cost to serve was clearly out of line with what people would pay. The best-funded company in the industry is making portfolio calls on unit economics. If your AI feature's inference bill grows faster than the revenue it drives, you are looking at the same decision.
$122 billion at $852 billion
OpenAI closed its round on March 31 with $122 billion committed at an $852 billion post-money valuation. Amazon put in $50 billion, $35 billion of which is contingent on an IPO or an AGI milestone. Nvidia and SoftBank put in $30 billion each. OpenAI also said it is generating about $2 billion a month in revenue, has more than 900 million weekly ChatGPT users, and gets over 40% of revenue from enterprise.
Disclosure: I work at AWS, so I have an obvious interest in the Amazon piece. The structure is worth reading either way. A tranche that pays out on an IPO tells you investors expect a listing, and soon.
Anthropic's next model leaks early
On March 26, a misconfigured CMS exposed internal Anthropic documents describing a model called Mythos, positioned above Opus. Anthropic confirmed it existed. The formal announcement came in April, and it became the story of the spring.
What March tells us
- Agents are the default workload. Computer use in the base model and inference-specific silicon point the same way.
- Unit economics decide what survives. Sora is the proof.
- The capital is heading to public markets. IPO-contingent tranches mean investors are planning on a listing.
The question for any AI product in 2026 is whether it earns more than it costs to run. OpenAI just answered it for one of its own.