Tech • AI • Robotics
Unitree Superman, $700B AI infra, Antimatter's compute bet
Briefing
Hyperscalers unleash $700B buildout
Amazon, Google, Microsoft and Meta are projected to spend more than $700 billion on AI infrastructure this year, roughly 4x the level of four years ago and nearly 80% above last year. The surge is making compute, memory, networking and model-serving operations a core engineering battleground rather than a back-office concern. McKinsey estimates nearly $7 trillion could flow into data centers by 2030, underscoring how capital intensity is reshaping the sector. The immediate beneficiaries are vendors and operators tied to GPUs, high-speed interconnects, storage and orchestration.
Antimatter pivots into compute and energy
Antimatter, launched in April 2026, is advancing a contrarian thesis that durable AI value may migrate away from proprietary models and toward infrastructure, routing and energy access. The company evolved out of Hivnet, moving from distributed storage into distributed computing after rejecting earlier blockchain-based storage architectures as commercially weak. Between 2022 and 2026, the strategy reportedly shifted several times, from software to compute capacity and then to power availability. The bet is that lower-cost, globally distributed infrastructure can compete where centralized AI supply remains constrained.
Unitree Superman targets robot speed records
Unitree unveiled its humanoid Superman with headline claims of a 2-meter standing vertical jump and a top speed of 12.66 meters per second. That velocity would slightly exceed Usain Bolt's peak recorded race speed of 12.42 meters per second, while the jump claim surpasses elite human standing benchmarks near 1.8 meters. With legs only about 0.85 meters long, the performance points to unusually powerful actuators and tightly tuned gait control. The release highlights how Chinese robotics groups are using athletic demonstrations to establish technical credibility ahead of commercial scale-up.
Shanghai IPO sharpens Unitree scrutiny
The Superman reveal arrived just three days before Unitree was due to begin trading in Shanghai, tying the demo directly to investor attention. The company raised 6.1 billion yuan, about $95 million, in its initial public offering. The timing underscores the financial premium now attached to humanoid robotics narratives, especially in China, where shipment leadership has become a strategic talking point. It also raises the usual question of how much spectacle translates into deployable autonomy and repeatable industrial use.
GPUs cement dominance in model serving
Modern AI systems are increasingly defined by the practical realities of serving large model files and executing massive parallel math. Typical model weights can range from roughly 2 GB for smaller systems to around 140 GB for a 70 billion-parameter model, with frontier models stretching into the hundreds of gigabytes. That footprint, combined with the need for billions of simultaneous multiplications, keeps GPUs decisively ahead of CPUs for inference workloads. The result is a stack optimized around accelerator clusters, memory bandwidth and low-latency model servers.
Kubernetes becomes AI operations backbone
As model-serving fleets scale, Kubernetes-based orchestration is emerging as the operational layer for scheduling, resiliency and resource efficiency. Engineering demand is shifting beyond model development toward infrastructure disciplines such as cluster management, GPU allocation and inference routing. That change reflects a broader maturation of the market, where uptime, cost control and deployment speed can matter as much as benchmark quality. In practice, AI competitiveness increasingly depends on the software that keeps accelerators utilized.
AWS blueprint digitizes factory defect tracking
A compact AWS smart-factory design shows how manufacturers can replace paper defect logs with near-real-time reporting using mostly serverless components. Workers submit a defect through a simple dashboard in about 10 seconds, after which API Gateway passes the request to Lambda, which creates an ID, timestamps the event and stores it in DynamoDB. Urgent cases can trigger immediate email notifications, while routine summaries are distributed on a schedule. The design emphasizes low operational overhead and fast adoption on the factory floor.
Serverless architecture cuts industrial friction
The factory stack uses a small EC2 instance with Nginx for the web front end, while browser requests flow through API Gateway with CORS enabled instead of connecting directly to the database. DynamoDB runs in on-demand mode to absorb variable traffic without capacity planning, and queue-based decoupling helps keep the workflow resilient during spikes. Architecturally, it is a straightforward pattern: minimal fixed infrastructure, event-driven processing and inexpensive scaling. That makes it a useful template for digitizing narrow industrial workflows without a large platform team.
Videos covered
Cloud, IA, Énergie : des Frenchies contre l’empire américain !
- •Investor trust over market validation
- •From distributed storage to AI infrastructure
- •A contrarian thesis on timing
AI Infrastructure Explained (GPUs, vLLM, and LLM-D)
- •Spending surge reshapes infrastructure demand
- •Models are software structures plus learned weights
- •Why laptops struggle and GPUs dominate
Un pays qui s'embrase, une caste qui s'embrasse... Avec Alexis Poulin
- •A village buried in minutes
- •Heavy human toll
- •A preventable industrial failure
China Just Dropped Superman - AI Robot With Superhuman Abilities
- •Superman’s headline performance
- •Control matters as much as power
- •An IPO-week launch with financial stakes
Build a Defect Tracker Using AWS (Serverless Project)
- •Factory-floor reporting in seconds
- •One server, two jobs
- •API Gateway as the public entry point