Tech Digest – March 17, 2026

The Infrastructure Arms Race

Nvidia Projects $1 Trillion in AI Chip Sales Through 2027 — and Builds the Coalition to Prove It

Jensen Huang announced that Nvidia’s AI processors are expected to generate $1 trillion in cumulative sales through 2027. To support that forecast, Nvidia launched the Nemotron Coalition, offering DGX Cloud compute to AI model partners — including Mistral, Perplexity, Cursor, and Black Forest Labs — who contribute data and expertise toward open models running on Nvidia silicon. On the hardware side, new liquid-cooled Vera CPU racks pack 256 custom processors designed specifically for reinforcement learning and agentic workloads — tasks that GPUs alone cannot efficiently handle.

Note: The Vera CPU is the detail that matters for planning: GPU infrastructure handles training, but agentic workloads — autonomous task execution, multi-step reasoning, workflow automation — require different silicon. Institutions evaluating AI tool deployments are increasingly dependent on infrastructure choices made three layers up the stack.

Sources: Bloomberg, Data Center Dynamics, The Register

Data Centers Overtake Offices in US Construction Spending — For the First Time

In December 2025, US construction spending on data centers exceeded spending on office buildings for the first time: $3.57 billion to $3.49 billion. The shift is expected to accelerate as AI automates functions that once filled those offices.

Note: The buildings follow the work. For public institutions planning long-term facility and infrastructure budgets, this reordering of construction priorities is a structural signal about where capital — and capacity — is flowing.

Sources: Bloomberg

Great Sky Claims Million-Fold Video Processing Speedup With Superconducting Light-Based Chip

Great Sky, a Boulder-based startup founded by former NIST researchers, has taped out its first chips based on a Superconducting Optoelectronic Network (SOEN) — an architecture that uses light rather than electrons to communicate data. The company claims its system can process video more than one million times faster than conventional GPU-based models, with lower energy consumption than a standard house. It raised $14 million in seed funding led by Bison Ventures. Cryogenic operation and scaling from prototype to production remain significant engineering challenges.

Note: The claim is extraordinary and unproven at scale. But the context matters: this is one of several simultaneous challenges to the GPU-centric AI stack — alongside neuromorphic chips, photonic processors, and alternative architectures. Institutions locked into multi-year AI infrastructure contracts should note that the hardware paradigm underneath those commitments is actively contested.

Sources: Semafor, Business Wire

Capital Commits to AI at Scale

Meta, OpenAI, and Private Equity Reshape How AI Infrastructure Gets Built — and Distributed

Meta has committed up to $27 billion over five years to Nebius for dedicated AI infrastructure. OpenAI, meanwhile, is pivoting Stargate — its high-profile data center initiative — away from building facilities toward renting cloud capacity, in order to accelerate deployment timelines. Separately, OpenAI is courting private equity firms including TPG, Advent, and Bain Capital for a $10 billion enterprise AI venture designed to distribute its products across portfolio companies and offer a lifeline to holdings facing AI-driven disruption.

Note: The private equity angle is structurally significant. AI adoption that travels through existing enterprise relationships — rather than requiring institutions to seek out new vendors — arrives without a visible procurement step. For public sector organizations working alongside private sector partners, this changes the timeline for AI exposure in their operational environment.

Sources: Bloomberg, The Information, Reuters

Agentic Tools Enter the Enterprise

OpenAI Pivots to Business Users as Codex Hits 2 Million Weekly Active Users — Up 4× Since January

OpenAI’s CEO of applications Fidji Simo told staff that the company is refocusing around coding and business users, citing Anthropic’s growth as a direct competitive signal. Codex, OpenAI’s coding assistant, now has 2 million weekly active users — nearly four times the figure from January — with API usage jumping 20% following the launch of GPT-5.4. Simo confirmed the company is building an enterprise deployment arm and expects more detail to follow shortly.

Note: 4× user growth in two months on a coding assistant is a procurement signal. Software development workflows inside public institutions — from internal tools to citizen-facing platforms — are the near-term surface where this lands. Institutions that haven’t yet inventoried how their development teams work should start.

Sources: Wall Street Journal, Fidji Simo / X

Nvidia Wraps OpenClaw Agent Platform With Privacy and Security Guardrails

NemoClaw is Nvidia’s implementation of the OpenClaw agentic platform, packaged with built-in privacy and security controls. The release positions compliance guardrails as a standard product feature — part of the agent deployment stack from the start, rather than a layer added after procurement.

Note: Guardrails as a product feature, not an integration task, meaningfully lowers the barrier for institutional adoption — and raises the expectation that vendors without them are behind the curve.

Sources: The New Stack

Regulated Sectors Go All-In

Roche Deploys 3,500 Blackwell GPUs for Biological Foundation Models and Drug Discovery

Pharmaceutical company Roche has deployed 3,500 Nvidia Blackwell GPUs across a hybrid cloud and on-premises environment to support biological foundation models and accelerate drug discovery research. The deployment is one of the largest AI infrastructure commitments by a regulated-sector institution to date.

Note: A heavily regulated institution — operating under strict data governance and compliance requirements — choosing hybrid cloud rather than full cloud deployment is a useful reference point. It demonstrates that compliance constraints and large-scale AI investment are not mutually exclusive.

Sources: Data Center Dynamics

Physics Simulations Run 400× Faster. The First Synthetic Bacterial Cell Is Alive. Research Timelines Are Compressing.

Two separate milestones this week signal the same underlying shift. The THOR AI framework from the University of New Mexico and Los Alamos National Laboratory used tensor networks to solve configurational integrals 400 times faster than advanced simulation methods. Separately, J. Craig Venter and collaborators demonstrated the first living synthetic bacterial cell, transplanting a complete genome into a dead cell — a landmark in synthetic biology that required computational tools to design and validate. Both outcomes represent research processes that previously took years being compressed into machine-executable loops.

Note: The institutional implication isn’t about biology or physics specifically — it’s about planning cycles. If research that used to take a decade is completing in months, the assumption that institutions have years to evaluate and adopt new capabilities is no longer reliable.

Sources: ScienceDaily / UNM, bioRxiv / Venter et al.

Energy and Governance Shifts

Sodium-Ion Batteries Hit the US Grid in a First-of-Its-Kind Pilot — Cutting Stored Energy Costs by ~50%

A first-of-its-kind sodium-ion battery pilot has gone live on the US Midwestern grid, with stored energy costs roughly half those of comparable lithium-ion installations. Sodium-ion technology avoids the lithium and cobalt supply dependencies that have driven up battery costs, and the technology is compatible with existing manufacturing infrastructure.

Note: Cheaper grid storage changes the economics of the AI buildout — and potentially the energy exposure of institutions managing their own infrastructure. For EU institutions tracking energy cost forecasts tied to data services and digital infrastructure, the direction of stored energy pricing is worth watching.

Sources: Electrek

SEC Moves to Eliminate Quarterly Earnings Reports in Favor of Semiannual Filing

The US Securities and Exchange Commission is preparing to eliminate mandatory quarterly earnings reporting, shifting public companies to semiannual disclosure. The stated goals are cost reduction and discouraging short-term decision-making. The change would represent the most significant shift to US financial reporting requirements in decades.

Note: The EU has its own reporting calendars and disclosure frameworks, but regulatory divergence between the US and EU on transparency timelines creates compliance planning complexity for any institution operating across both environments. Whether European regulators follow or maintain quarterly requirements is a question worth tracking now.

Sources: Reuters