Tech • AI • Robotics
AI creates viruses, devours resources, goes into space... and France is losing!
9.3/10TL;DR
AI, custom chips and power constraints are converging into a new phase of competition in which breakthroughs in biology and computing are advancing faster than governance and infrastructure.
KEY POINTS
AI-designed viruses reach the lab
Researchers using Evo 2, a genome model developed with Nvidia and major academic labs including Stanford, generated full viral genomes that did not previously exist in nature. After screening hundreds of candidates and synthesizing 285 in the lab, scientists identified 16 viable bacteriophages able to infect Escherichia coli. The work marks a major milestone: an AI system produced complete genomes that became functioning biological entities once synthesized.
Medical promise comes with biosecurity risk
The new viruses target bacteria rather than human cells and were constrained to a narrow biological domain related to phiX174 and E. coli. That makes the result less alarming than a human pathogen scenario, but it still demonstrates that AI can design living systems from genomic patterns. The same capability could help fight antibiotic resistance, since a cocktail of multiple phages can make bacterial adaptation much harder.
Open access raises governance questions
The concern is not only scientific progress but accessibility. If genome-generating systems become broadly usable through public interfaces or replicated models, a capable lab could potentially synthesize outputs in the real world. That prospect is intensifying calls for international rules on what kinds of biological design AI should and should not enable.
OpenAI unveils its own inference chip
OpenAI has introduced an inference-focused chip called Jalapeño, designed with Broadcom to reduce dependence on external hardware suppliers. Unlike training accelerators, the chip is built for serving models efficiently, combining compute and memory functions for faster response generation. The move places OpenAI more directly in competition with Nvidia and other chip leaders.
Performance gains target speed and power use
Early claims suggest Jalapeño can deliver substantially better throughput than current top-tier systems on selected open models, with efficiency gains of roughly 2x to 4x for the same energy use. That matters because inference is becoming one of the biggest recurring costs in AI. Even if first deployments remain limited through late 2026 and larger rollout slips into 2027, the strategic message is clear: major AI firms want control over both software and silicon.
Apple pushes local AI with new chips
Apple has also expanded its hardware push with new M6 and M5 Ultra chips aimed at powerful on-device inference. The approach differs sharply from cloud-centric rivals: rather than concentrating AI in hyperscale data centers, Apple is strengthening the ability to run large models locally on high-end machines. Systems with up to 512 GB of unified memory could make advanced local inference practical for developers and enterprises, albeit at premium prices.
Electricity becomes the limiting factor
The rapid advance of AI hardware is colliding with a harder constraint: power supply. In the United States, utilities and policymakers are increasingly forced to weigh demand from data centers against households, hospitals and schools. The strain is strongest in areas already saturated with computing infrastructure, where AI growth is outpacing available clean and reliable generation.
US data centers fuel a return to gas
Industry data indicate a renewed surge in gas turbine demand as developers seek the fastest way to power AI facilities. Orders are reportedly stretching years ahead, and large operators including Meta and projects linked to Elon Musk are associated with aggressive efforts to secure generation capacity. Critics warn that this deepens pollution, raises local opposition and shifts costs onto residents through higher power prices.
Power prices spotlight the imbalance
In parts of the US market such as Virginia, wholesale electricity prices linked to heavy data-center demand are far above levels commonly seen in France, with figures cited around $600 per MWh and possible spikes toward $1,000. Grid operators are increasingly considering rules that would require new data centers to bring their own generation capacity before connecting. The debate is no longer abstract: AI expansion is directly shaping energy policy and consumer bills.
Hugging Face becomes a takeover target
Hugging Face, founded by French entrepreneurs and now central to the global ecosystem for open models, datasets and AI tools, is reportedly being discussed in a deal around $13 billion. The company was valued at $4.5 billion after a $235 million fundraising round in 2023 and has presented itself as nearing profitability. Its rise reflects the growing strategic value of open AI infrastructure rather than heavy spending alone.
A French-origin open champion faces a crossroads
The company has built influence by hosting and distributing much of the world’s open AI ecosystem while trying to avoid capture by any single dominant lab. It previously resisted offers that could have tied it more closely to a major platform player. A sale to a leading AI company could transform or effectively end its role as a neutral open hub, while a purchase by a broader software or financial group would still mark the loss of a rare French-origin strategic asset.
CONCLUSION
The latest advances show AI moving simultaneously into biology, hardware and energy systems. The central challenge is no longer invention alone, but whether public safeguards and strategic choices can keep pace with technologies that are becoming foundational infrastructure.
Explain this