Enterprise AI has spent two years justifying itself. The results arriving this week suggest it no longer needs to: operators are converting experiments to contracts, infrastructure is being reset around singular competitive bets, and the physical world is starting to come within reach. Four developments frame where things stand.

Palantir's Numbers Settle the Debate on Enterprise AI Revenue

Palantir reported Q2 2026 revenue of $1.94 billion, up 93% year on year. US commercial revenue reached $764 million, a 149% year-on-year increase. Adjusted operating income hit $1.19 billion at a 62% margin; GAAP operating income was $912 million at a 47% margin. The company's Rule of 40 score reached 155 — a level that essentially does not exist in enterprise software growing at this rate. Full-year revenue guidance was raised to $8.15–8.16 billion, implying 82% growth for the year. The stock rose 12.5% in after-hours trading.

For an operator evaluating whether to deploy AI infrastructure seriously, the Palantir result answers the most common board-level objection: that AI investment does not generate measurable return. Palantir's AIP product is the layer that sits between raw model output and operational decisions; its acceleration in US commercial revenue reflects enterprises committing to that layer, not just experimenting with chat interfaces. Management described demand as having been "unleashed" around AI sovereignty — a phrase that captures something real: procurement decisions that were stalled are now moving.

The read-through is not Palantir-specific. It signals that the budgets allocated to AI pilots over the past two years are converting to multi-year platform contracts. Operators who are still treating AI as a cost centre experiment should note the pace at which their peers are not.

Amazon Resets Around One Frontier Model, Ending Its Nova Family Strategy

Amazon announced on July 28 that it is winding down active development of four of its flagship Nova AI models — Premier, Omni, Reel, and Canvas — placing them in maintenance mode with end-of-life dates in September 2026. Resources are shifting to Frontier Model Research (FMR), a new internal effort led by Pieter Abbeel, a UC Berkeley professor who joined Amazon through its 2024 acquisition of robotics startup Covariant. A new flagship foundation model is expected at re:Invent later in 2026. The Nova name may be retained.

Amazon is keeping Nova 2 Lite and Nova Sonic as functional models and retaining Nova Forge and Nova Act for customisation and agent workflows. The departure of Rohit Prasad, who oversaw the Nova programme, accompanies the reorganisation.

The signal for operators is structural. Amazon is the third major AI provider, after OpenAI and Google, to reorganise its effort around one competitive frontier model rather than a family of models calibrated to different price points. That convergence changes vendor selection: you are no longer choosing a product tier, you are choosing a lab and its trajectory. It also reduces the operational overhead of tracking which model version applies to which workload — but it raises the stakes of the choice considerably.

Google DeepMind Releases Gemini Robotics 2: Whole-Body Humanoid Intelligence

On July 30, Google DeepMind released Gemini Robotics 2, a family of models designed to control complete humanoid robot bodies. The system differs from its predecessor in a key respect: where the earlier model controlled primarily the upper body, Gemini Robotics 2 drives the entire robot — legs, torso, arms, and five-fingered hands — within a single model framework.

The release comprises three components:

  • A vision-language-action (VLA) model that converts visual and language inputs into physical movements in real time.
  • Gemini Robotics ER 2, an embodied reasoning model for multi-step task planning and multi-robot coordination.
  • Gemini Robotics On-Device 2, an optimised variant that runs locally on robotic hardware without cloud connectivity.

Demonstrations were conducted on Apptronik's Apollo 2 humanoid, which performed tasks including screwing in lightbulbs, sealing bags, tying trash bags, and walking across uneven terrain.

For operators in logistics, manufacturing, or facilities management, the transition from single-limb control to whole-body autonomy is the prerequisite for a robot to do genuine work in an unstructured environment. The on-device variant — which eliminates cloud-inference latency and the connectivity dependency — is the detail that matters most for production deployment at scale. The capability gap that separated demonstration from deployment has narrowed measurably.

Fifty Nations Convene in Geneva to Draft the First Sovereign AI Compact

From 12 to 14 August, more than 50 nations from the Global South will gather at the Palais des Nations in Geneva for the AI for Developing Countries Forum (AIFOD) summit, under the heading "Small Takes the Lead." The centrepiece is the Geneva Compact on AI Sovereignty — a framework document intended to codify principles of AI equity and autonomy for the majority of the world. Day three is dedicated to drafting and signing the first bilateral pooled-compute agreements: arrangements under which nations combine purchasing power to access computing infrastructure collectively, replicating at the infrastructure level what open-weight model releases have done at the software level.

The four working tracks cover access for small nations, small enterprises, small (locally runnable) models, and small languages — the dimensions where developing countries are most underserved by the current AI ecosystem. A follow-on convening is committed for Nairobi in 2027.

For a multinational operator, the summit is early signal for regulatory and procurement risk. Nations that formalise sovereign AI frameworks will begin setting their own data-localisation standards, model-audit requirements, and procurement conditions. The timeline for those rules is 2027, not 2030. Operators with exposure across African, South Asian, or Latin American markets should be tracking the Geneva Compact from its first draft.

The thread across these four developments is that AI is no longer being evaluated — it is being decided. Palantir's numbers confirm the revenue case at enterprise scale. Amazon's reset signals that the model-family era is giving way to singular bets. Gemini Robotics 2 puts the physical world within scope of the same model-driven logic. And the Geneva summit signals that the nations which had no seat at the AI table are beginning to build their own. Operators who are still at the evaluation stage are increasingly the exception, not the rule.