Uber is rolling ChatGPT Work out to more than 30,000 employees globally, expanding from a 1,000-person pilot and making it the company’s first organization-wide deployment of an agentic platform for knowledge work.
The tool connects to Google Drive, Gmail, Google Calendar, and Slack to research information, prepare meeting briefs, create presentations, analyze data, and complete multi-step workflows. GPT-6.1 Sol will be the default model, with voice and image generation also available.
The pilot found that roughly 80% of early users were non-engineers. Uber has not disclosed its spending or fixed use cases; it plans to track adoption and outcomes. Software teams will continue using Codex.
Why it matters: Uber gives OpenAI a 30,000-plus-employee proving ground for agentic work, moving the product beyond small-team experimentation. The rollout also gives Anthropic a clear enterprise benchmark: whether an AI agent can produce measurable workflow gains across a global company, not just assist individual users.
NVIDIA reports more than $300 billion in backlogged equipment and over 200 gigawatts of U.S. projects waiting in interconnection queues. Its DSX platform connects grid, facility and compute planning; Phaidra says its integration delivered about 10% more compute within the same power envelope.
U.S. Hunger uses Amazon Quick so leaders can explore organizational data without asking technology staff for a new pull. The nonprofit reports roughly 500% growth in four years versus 15–20% headcount growth, and uses Quick Spaces to align financial analysis, KPIs and program data.
Both examples replace siloed workflows with shared operating environments, but the performance gains remain vendor or customer claims, not independent benchmarks.
Why it matters: The operating advantage differs by group: NVIDIA customers can test power, cooling and compute before breaking ground, while U.S. Hunger’s 17 Quick users can follow questions without waiting for technology staff to build another data pull.
AWS made SageMaker AI serverless model customization training sessions generally available, letting companies write custom post-training loops without managing GPU infrastructure. Customers control loss functions, reward signals, and training stages; AWS handles GPU allocation, distributed execution, fault recovery, and checkpoint storage.
The sessions support SFT, DPO, RLVR, RLAIF, and multi-turn reinforcement learning, with either LoRA or full fine-tuning. Billing is token-based: the meter pauses while a customer’s reward function scores a batch, and AWS says capacity reservations are unnecessary.
ImagineSoftware is using the service for healthcare claim-denial triage; Karini AI fine-tuned gpt-oss-20b for multi-turn tool calling against an Amazon Neptune knowledge graph.
Why it matters: For teams like ImagineSoftware and Karini AI, custom reinforcement learning no longer requires choosing instance types, managing clusters, or writing distributed-training configuration; AWS supplies that infrastructure while customers retain control of the training loop.
Microsoft’s MAI-DxO diagnosed complex medical cases correctly 83% of the time, versus 36% for primary care doctors and general internists. The comparison came from a simulation using cases from the New England Journal of Medicine, not live patient care; specialists and physician teams were excluded.
MAI-DxO coordinates AI agents that question patients, order tests, and debate diagnoses. Its strongest result came with OpenAI’s GPT-5.5, lifting accuracy from 72% to nearly 87% on complex cases. The gains disappeared in emergency-department cases, where both versions scored about 85% at similar cost.
Microsoft will test DxO alongside standard care at Beth Israel Deaconess Medical Center. Mustafa Suleyman also said a Mayo Clinic healthcare foundation model is planned for release next year.
Why it matters: For Microsoft, the decisive milestone is not the 83% simulation score but the Beth Israel evaluation, which will show whether MAI-DxO can survive real emergency and urgent-care workflows. Until then, the result is evidence of model orchestration—not proof that hospitals can replace specialist judgment.
Google’s Agent Compute Engine gives every AI agent its own isolated virtual machine, and the service already powers Gemini. Google plans a preview in coming months to cut the infrastructure cost of running agentic systems.
Each microVM runs on Google’s Titanium chip, pauses when idle, and shares memory with other VMs. Google says that design fits 10 times as many agents on one server.
The open-source Agent Substrate Kubernetes runtime uses the same architecture. In a 200,000-agent test, suspend-and-resume took under 500 milliseconds. Google is also adding token tracking, model routing, and identity support from Ping Identity and SailPoint across Cloud Run, GKE, and managed endpoints.
Why it matters: For Cloud Run and GKE customers, Google’s pitch is economic rather than model-focused: if its 10x density claim holds in real workloads, teams could run a given agent load with less server capacity.
EXL’s Engineering the Harness series argues that enterprise AI’s advantage increasingly comes from the system around a model: memory management, scheduling and inference optimization. That system determines whether a model reaches production reliably, economically and at scale.
Proof-of-concept teams often prioritize accuracy, but production systems must handle latency, concurrency, memory pressure, GPU utilization, infrastructure cost and unpredictable traffic. The key-value cache grows with active requests and context length, potentially exhausting usable GPU memory even when compute capacity remains available.
EXL recommends workload-specific architecture, clear service-level objectives and iterative measurement. Quantization, speculative decoding and optimized attention should target the measured bottleneck, allowing model changes to become configuration decisions rather than platform rebuilds.
Why it matters: For technical executives applying EXL’s framework, a model swap can become a configuration decision instead of an architectural reset—if the serving system meets workload-specific latency, concurrency and cost constraints.
Amazon SageMaker AI now lets researchers control an LLM training loop without provisioning or operating the GPUs underneath it. Developers call operations such as forward_backward(), optim_step(), and sample() from a Python process while SageMaker keeps the model loaded on managed accelerators and handles distributed execution, fault recovery, and checkpoint storage.
The APIs target custom supervised fine-tuning, preference optimization, and reinforcement learning workflows that standard recipes cannot express. SageMaker bills usage per token processed; reward logic running on the developer’s machine consumes no GPU time. AWS also released a cookbook with ready-to-run SFT and RL recipes.
Why it matters: Researchers building custom SFT and RL algorithms can iterate on losses, rewards, and training steps without taking on GPU allocation, distributed execution, or fault recovery, while SageMaker charges for the compute their loops actually consume.