# Infrastructure, Open Models, and Agent Workflows Define the Day

*By AI News Digest • March 12, 2026*

Sam Altman used BlackRock's infrastructure summit to argue that frontier AI now depends as much on power, construction, and inference economics as on model progress. Elsewhere, NVIDIA launched a major open model for agentic systems, enterprise tools kept shifting toward orchestrated digital work, and governance proposals became more concrete.

## Infrastructure became the main story

The clearest pattern today was that frontier AI is being described in terms of **power, chips, and construction** as much as model intelligence [^1].

### OpenAI framed frontier progress as a buildout problem

At BlackRock's US Infrastructure Summit, Sam Altman said OpenAI is already training at its first Stargate site in Abilene and described the challenges of getting gigawatt-scale campuses running, from unexpected weather to supply-chain issues and the need for many organizations to work together under pressure [^1][^2]. He also said OpenAI's new partnership with the North American Building Trades Unions reflects a practical constraint: AI growth depends on physical infrastructure such as power plants, transmission, data centers, and transformers, plus more skilled trades workers to build them [^1][^2].

*Why it matters:* The bottlenecks around frontier AI are increasingly physical, not just algorithmic.

### Altman said costs are falling fast — and specialized inference hardware matters more

Altman said OpenAI's first reasoning model, o1, arrived about 16 months ago, and that getting the same answer to a hard problem from o1 to GPT-5.4 now costs about **1,000x less** [^1][^2]. He also said the company is building an **inference-only** chip optimized for low cost and power efficiency, with first chips expected to be deployed at scale by year-end [^1][^3]. Altman added that the past few months marked a threshold of major economic utility for these systems, especially in coding and other knowledge work [^1].

> "To get the same answer to a hard problem from that first model to 5.4 has been a reduction in cost of about a thousand X." [^1]

*Why it matters:* Capability gains are now being paired with meaningful cost compression, which is what turns impressive demos into deployable systems.

## Open models and agent products widened the deployment race

### NVIDIA released an open model aimed squarely at agentic AI

NVIDIA launched **Nemotron 3 Super**, a 120B-parameter open model with 12B active parameters, a **1-million-token context window**, and high-accuracy tool calling for complex agent workflows [^4]. NVIDIA said it delivers up to **5x higher throughput** and up to **2x higher accuracy** than the previous Nemotron Super model, and is releasing it with open weights under a permissive license for deployment from on-prem systems to the cloud [^4].

*Why it matters:* This is a substantial open-model push focused on enterprise-grade agents, not just model openness as a slogan.

### Enterprise products kept moving from chat toward orchestrated work

Perplexity launched **Computer for Enterprise**, saying it can run multi-step workflows across research, coding, design, and deployment by routing work across **20 specialized models** and connecting to **400+ applications** [^5]. The company said its internal Slack deployment performed **3.25 years of work** and saved **$1.6M** in four weeks, and that it is now exposing some of the same orchestration through a model-agnostic API platform [^6][^7][^8].

The same shift appeared elsewhere: Replit introduced **Agent 4** for collaborative app-building with an infinite canvas and parallel agents [^9], while Andrej Karpathy argued this does not end the IDE so much as expand it into an **"agent command center"** for managing teams of agents [^10][^11].

*Why it matters:* A growing set of products is treating AI less like a single assistant and more like a coordinated workforce.

## Governance ideas got more operational

### Anthropic created a new public-benefit function around powerful AI

Anthropic said Jack Clark is becoming **Head of Public Benefit** and launching **The Anthropic Institute** to generate and share information about the societal, economic, and security effects of powerful AI systems [^12][^13][^14]. Anthropic said the institute will bring together machine learning engineers, economists, and social scientists, using the vantage point of a frontier lab to inform public understanding [^15][^16].

*Why it matters:* Frontier labs are starting to formalize impact analysis as an institutional function, not just a policy sideline.

### A biosecurity proposal focused on restricting dangerous data, not shutting down open science

Johns Hopkins researcher Jassi Pannu outlined a **Biosecurity Data Level** framework that would keep roughly **99%** of biological data open while adding controls only to the narrow slice of functional data that links pathogens to dangerous properties such as transmissibility, virulence, and immune evasion [^17]. She also pointed to model-holdout results suggesting that removing human-infecting virus data can sharply reduce dangerous biological capabilities while leaving desirable capabilities intact [^17].

*Why it matters:* It is one of the clearest middle-ground governance proposals on the table: preserve open research broadly, but treat the most dangerous capability-enabling data as a controlled resource.

---

### Sources

[^1]: [DIRECTO | Sam Altman participa en la Cumbre de BlackRock sobre Infraestructura de EE UU | ELPAÍS](https://www.youtube.com/watch?v=Y74oyV6oSGo)
[^2]: [LIVE: Sam Altman speaks at BlackRock's US Infrastructure Summit](https://www.youtube.com/watch?v=Wc3OSJ2_6CA)
[^3]: [WATCH LIVE: OpenAI CEO Sam Altman speaks at BlackRock's US Infrastructure Summit](https://www.youtube.com/watch?v=nGjTMBb2dSE)
[^4]: [New NVIDIA Nemotron 3 Super Delivers 5x Higher Throughput for Agentic AI](https://blogs.nvidia.com/blog/nemotron-3-super-agentic-ai)
[^5]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2031799033489211771)
[^6]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2031796262010593635)
[^7]: [𝕏 post by @perplexity_ai](https://x.com/perplexity_ai/status/2031828396435771563)
[^8]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2031829586133332176)
[^9]: [𝕏 post by @amasad](https://x.com/amasad/status/2031755113694679094)
[^10]: [𝕏 post by @karpathy](https://x.com/karpathy/status/2031767720933634100)
[^11]: [𝕏 post by @karpathy](https://x.com/karpathy/status/2031616709560610993)
[^12]: [𝕏 post by @jackclarkSF](https://x.com/jackclarkSF/status/2031746605117010245)
[^13]: [𝕏 post by @jackclarkSF](https://x.com/jackclarkSF/status/2031746606496944609)
[^14]: [𝕏 post by @jackclarkSF](https://x.com/jackclarkSF/status/2031746607994245278)
[^15]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2031674092290474421)
[^16]: [𝕏 post by @AnthropicAI](https://x.com/AnthropicAI/status/2031674090604417307)
[^17]: [Bioinfohazards: Jassi Pannu on Controlling Dangerous Data from which AI Models Learn](https://www.cognitiverevolution.ai/bioinfohazards-jassi-pannu-on-controlling-dangerous-data-from-which-ai-models-learn)