Wall Street has questions about the price of AI

Good day!

AI spending attracted plenty of attention this week, and not all of it was positive.

Meta raised the lower end of its annual capital expenditure forecast to $135 billion, with the upper end still at $145 billion. A day later, the Nasdaq fell 2.6% as investors became increasingly nervous about the enormous sums going into AI infrastructure.

There were some interesting developments beyond the financial markets, too. Microsoft published research showing impressive results for an AI diagnostic system, Walmart outlined how customers could authorise AI assistants to make purchases, and Canva launched its first image-generation model.

Meanwhile, the discovery and disclosure of critical security vulnerabilities using AI triggered an international scramble to protect websites, including government systems.

Microsoft published research in Science examining an AI system designed to work through medical diagnoses step by step.

Its Microsoft AI Diagnostic Orchestrator, known as MAI-DxO, uses multiple AI systems to consider possible diagnoses, request information, evaluate tests and refine its conclusions.

In a controlled assessment involving complex medical cases, the system achieved 83% diagnostic accuracy, compared with 36% for generalist physicians using reference materials. On a separate set of simulated emergency-department cases, it scored 80%, compared with 47% for the physicians.

Those numbers are remarkable, but they need context. The evaluations used historical, text-based cases and did not establish that the system outperforms specialist medical teams or delivers better outcomes in clinical practice.

Microsoft is now preparing real-world evaluations with hospital partners. The next important results will concern whether the system improves clinical decision-making, not simply how well it performs against a benchmark.

Thursday brought one of the sharpest technology sell-offs in months.

The Nasdaq fell 2.6%, its largest one-day percentage decline since early June. The Philadelphia Semiconductor Index dropped 6.7%, while Meta shares declined approximately 7% despite its better-than-expected quarterly revenue.

The immediate concern was the financial pressure created by the AI investment cycle.

Investors have generally welcomed evidence that companies are building more computing capacity. But the scale of expenditure is making free cash flow, financing requirements and future returns harder to overlook.

Legal technology company Legora announced a $550 million Series E funding round, valuing the business at $8.8 billion.

The company says its annual recurring revenue has exceeded $200 million, having reached the $100 million milestone less than six months earlier.

Those are substantial figures for a relatively young enterprise software business operating in an industry traditionally associated with lengthy procurement processes and cautious technology adoption.

Legora provides AI software for legal research, document review and complex legal workflows. Its customer base includes law firms and in-house legal departments, with the US now representing its largest market by revenue.

The funding is another indication that investors see significant commercial potential in specialised AI products aimed at professions with high-value, information-intensive work.

Walmart announced a new framework designed to let customers use AI assistants to purchase products on their behalf.

Its proposed Agent Profiles will allow customers to give individual AI agents their own permissions, track purchases separately and decide whether every transaction requires approval.

Customers will also be able to establish spending limits, such as $50 or $100, and revoke an agent's access.

Walmart plans to launch Agent Profiles over the coming month and is developing an Agent Gateway to help third-party AI applications interact with its shopping systems.

The commercial significance extends beyond a more convenient checkout process.

If consumers increasingly ask AI assistants to research and purchase products, retailers will need to think about how their catalogues, prices, inventory and customer relationships work when software performs part of the shopping journey.

Walmart appears interested in making its products accessible to those systems while retaining control over the transaction.

Amazon is helping fund development work for a proposed nuclear power project in Indiana, alongside X-energy and American Electric Power.

The plan involves 12 small modular reactors, potentially delivering approximately 960 megawatts of electricity near the site of a former coal-fired power station.

It is still at the feasibility stage. The reactors have not been built, and the project will require substantial further development and regulatory approval.

Nevertheless, it provides another concrete example of how growing electricity requirements are influencing technology companies' energy investments.

AI infrastructure needs reliable power, and companies are exploring agreements that extend well beyond conventional renewable electricity purchases.

For Amazon, the arrangement is part of a broader effort to support the development of new nuclear generating capacity.

For the energy sector, the increasing involvement of technology companies could become an important source of capital for projects that have historically struggled with long development periods and high upfront costs.

Anthropic announced Claude Duet, a standalone AI application designed to help users manage information and complete tasks across connected services.

Rather than relying entirely on a user to initiate each conversation, Duet is designed to provide morning briefings, identify items requiring attention and continue working on assigned tasks.

It can help with activities ranging from scheduling and administrative work to research, forecasts and software development.

Anthropic says the assistant seeks approval before taking actions on a user's behalf.

For now, this is a research preview available through a waitlist, initially aimed at US subscribers on Claude Pro and Max plans.

The announcement raises an interesting product question: how much initiative will users actually want from their AI assistants?

Proactive notifications and task management could be useful. But they will also need to be sufficiently accurate, relevant and restrained to avoid creating another source of interruptions.

Canva introduced Canva Craft, its first image-generation foundation model developed by the company's own research team.

The model can generate images at up to 2K resolution and is available through Canva and its developer API.

Canva says Craft has been designed specifically for creative production, with an emphasis on composition, text rendering and outputs that can be refined within existing design workflows.

It also reported strong results in crowdsourced image-generation rankings, although those performance claims should be considered alongside independent testing and real-world user experience.

The decision to develop its own model is commercially interesting.

Canva has already built a sizeable distribution platform for creative software. Owning more of the underlying AI technology could give it greater control over product quality, development priorities and the economics of delivering image-generation features.

It also adds another competitor to a market already crowded with dedicated image-generation companies.

Google Cloud announced that OpenAI's latest models are becoming available through its Model Garden platform.

The initial selection includes GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, covering more demanding reasoning tasks through to lower-cost, high-volume applications.

Access is initially limited to selected enterprise customers through a private preview.

This is a notable distribution arrangement because Google is also one of OpenAI's major competitors in AI model development.

For enterprise customers, the appeal is straightforward: access to models from competing developers through an existing cloud platform, with consolidated billing and familiar security controls.

For the providers, it reflects the growing importance of distribution and enterprise relationships alongside raw model performance.

The companies building models increasingly compete on capability while also finding commercial reasons to make those models available through rival platforms.

O T H E R_A I N E W S

Amazon Web Services introduced TorchNeuron for PyTorch developers, allowing existing PyTorch code to run on AWS Trainium AI chips with minimal changes. The software is intended to reduce the effort required to adopt Amazon's alternative to established AI computing hardware.

Google started bringing personalised Gemini assistance to cars, allowing eligible users to draw on information from connected Google services while driving. The opt-in rollout began with Android Auto, although the personalised feature is not currently available in the UK or European Economic Area.

Nvidia secured a new customer for its Vera central processors, with AI infrastructure startup Modal planning to adopt the chips alongside its existing Intel and AMD systems. Modal said early tests produced twice the agent-workload throughput of its current processors.

1Password launched 1Password for Agents, a platform designed to let authorised AI assistants sign in to services and complete approved tasks without directly exposing or storing users' passwords and credentials.

OpenAI acquired Orchestra Bio, a platform for coordinating complex drug-development programmes. The company plans to combine Orchestra's technology with Rosalind, its existing AI research tools for life sciences. Financial terms were not disclosed.

Otter.ai secured $320 million in growth financing from General Catalyst's Customer Value Fund. The non-equity financing will support enterprise expansion, international growth and further development of its AI meeting products.

Recursive released Flint, a coding model designed to run on laptops. The startup says AI played a substantial role in developing the model itself, making the release an interesting experiment in AI-assisted model development.

Nebius secured a $3.5 billion financing facility from Citi to support data-centre expansion and GPU purchases. The 364-day facility illustrates the substantial financing requirements facing specialist AI cloud infrastructure providers.

Anthropic faced new arguments in a major AI copyright case, as music publishers challenged its use of copyrighted song lyrics to train Claude. A California judge heard arguments concerning fair use but did not issue a ruling.

Warp Sciences launched with $215 million in funding, according to founder and former OpenAI science executive Kevin Weil. The startup aims to combine AI and robotics to automate laboratory experiments and accelerate drug discovery.

Thanks for reading, and see you next week!

The News Digest AI Team