The race now spans four connected layers

AI competition is no longer adequately described as a model leaderboard. The current delivery stack runs from model platforms and custom silicon through powered data-center capacity to robots operating in physical environments. TechReadly’s expanded edition follows 40 company developments across those four layers.

The companies are also crossing traditional boundaries. Apple and Amazon are pairing consumer or cloud distribution with model and silicon work. Meta is developing models, accelerators, and long-term energy supply. Chinese groups including Alibaba, Tencent, ByteDance, Baidu, Huawei, and Xiaomi are similarly connecting models to cloud services, devices, or robotics.

An announcement at one layer is therefore most useful when read against its dependencies. A stronger model still needs economical inference; a new accelerator still needs a qualified rack; a planned campus still needs power and commissioning; a robot still needs reliable hardware, data, and control software.

Delivery checkpoints matter more than headline scale

The clearest signal across the edition is the difference between an announced product, a system in production, a running service, and future capacity. NVIDIA says Vera Rubin racks are running at named cloud providers; AMD says Helios is in production; HUMAIN says its first Saudi system is serving customers. Those are distinct milestones, not interchangeable proof of broad availability.

The same discipline applies to data centers. OpenAI has broken ground on a planned 1 GW Michigan campus, while Amazon, Google, Meta, CoreWeave, and others have described new regions, leases, energy centers, or power agreements. Construction, grid connection, commissioning, and customer access remain the checkpoints that convert those plans into usable compute.

Robotics is the downstream integration test

Robotics makes the integration problem visible. US companies including Figure, Boston Dynamics, and Agility Robotics are moving from research demonstrations toward data collection, product programs, facilities, and customer deployments. Chinese companies including AGIBOT, Unitree, Xiaomi Robotics, and LimX Dynamics are broadening hardware portfolios and opening parts of their model or development stack.

None of those moves establishes that general-purpose humanoids are ready for broad deployment. They do create measurable next steps: customer task scope, fleet size, uptime, safety validation, external developer access, and repeatable performance outside curated demonstrations.

The practical competitive unit is becoming the delivered system. Model quality remains essential, but sustained advantage increasingly depends on coordinating silicon, memory, networking, power, cooling, software, data, mechanics, and operations.

What to watch next

  • Which model releases become durable products with clear pricing, access, and customer adoption.
  • When announced chips and racks become generally available capacity with independent workload results.
  • How much planned power and data-center capacity reaches construction, commissioning, and operation.
  • Whether robot programs disclose repeatable customer tasks, fleet size, uptime, and safety performance.
  • Where cross-layer control produces a real cost, distribution, or execution advantage.