94.5% Accurate.
Under 15 Seconds.

Swine farming operates under a persistent pressure: disease moves fast through a herd, veterinarians are unevenly distributed, and the gap between a sick pig and a confirmed diagnosis can cost producers significantly. Researchers at AXONS, working with Charoen Pokphand Foods, built a multi-agent AI diagnostic system trained on veterinary knowledge and capable of identifying swine diseases from clinical symptoms alone.

The system achieved 94.5% accuracy on disease diagnosis tasks, classified user queries correctly 95.23% of the time, and returned responses within 15 seconds on average. These are not benchmark numbers in isolation: the researchers compared multiple leading language models against each other and traced where the gains come from.

The architecture is a three-stage pipeline. An initial classifier routes each incoming query into one of four categories: knowledge retrieval, symptom-based diagnostic, queries requiring clarification, and general questions. When a symptom-based query is identified, a second stage collects symptoms in layers, beginning with general health indicators, then external signs, then specific symptom clusters tied to particular disease groups.

Once the symptom picture is assembled, multiple specialized agents analyze the data simultaneously. Each generates a confidence score for candidate diseases. Those scores are combined through a weighted fusion mechanism. A confidence threshold determines which diseases rise to the top, ranked from very high to low certainty. The third stage generates treatment recommendations using Retrieval-Augmented Generation, drawing from a domain-specific veterinary knowledge base rather than relying on model training alone.

The model comparison results are instructive. GPT-4o achieved 90.63% test accuracy with an 18.78-second average response. Gemini-1.5-Pro-002 reached 94.23% validation accuracy but dropped to 87.50% on the test set, a gap suggesting it may overfit to particular query patterns. The o1-mini model lagged on both accuracy and speed, averaging 29.38 seconds, which the researchers flag as a practical constraint for real deployment. These differences illustrate why the multi-agent design matters: pooling predictions across models with different strengths reduces the risk that any single model's failure modes drive the final output.

A farmer who notices reduced appetite, labored breathing, or skin lesions in the early morning often cannot reach a veterinarian until hours later. An AI system capable of asking structured diagnostic questions and returning a high-confidence provisional diagnosis within 15 seconds changes that timeline. In regions where veterinary access is severely limited, the system could serve as a primary triage mechanism, helping producers distinguish between conditions requiring immediate isolation and those that can be monitored.

American data centers consumed 183 terawatt-hours of electricity in 2024, more than 4% of national consumption. Analysts expect that share to exceed 9% by 2030. At the same time, the existing distribution network operates at roughly 40 to 45% utilization on average. More than half its capacity sits idle most of the time. Building new infrastructure to close the gap takes years. A 100-megawatt data center typically runs upward of $15 million per megawatt to build, with a three-to-five-year construction timeline.

SPAN's answer is XFRA: an outdoor compute node paired with a smart electrical panel and a whole-home battery, installed in residential and small commercial buildings. Software routes inference jobs across the distributed fleet based on available power and latency requirements. The cost comparison is material. Reaching 100 megawatts using XFRA requires installation in roughly 8,000 homes over about six months, at approximately $3 million per megawatt, compared to $15 million per megawatt for a conventional data center build.

NVIDIA's contribution is specific. The RTX PRO 6000 Blackwell Server Edition GPU selected for XFRA carries 96GB of GDDR7 memory, fifth-generation Tensor Cores with FP4 precision support, and can be partitioned into up to four isolated instances. Benchmarks indicate it delivers up to five times the LLM inference throughput of the previous-generation L40S chip, with roughly twice the price-performance of an H100 system for inference tasks. The GPU is passively cooled and rated for 24/7 operation, which matters when the hardware lives in a residential setting.

SPAN's approach differs from decentralized compute networks in two ways. First, XFRA uses uniform, enterprise-grade hardware rather than aggregating whatever happens to be available. Second, SPAN's orchestration software controls power at the circuit level through its smart panel integration, giving it scheduling capability a pure software marketplace cannot replicate. The company is not asking homeowners to contribute idle gaming rigs. It is installing standardized nodes and managing them like distributed utility assets.

A pilot deployment begins later this year across 100 newly constructed homes, representing about 1.25 megawatts of compute capacity and 1,600 liquid-cooled inference GPUs. PulteGroup, one of the largest U.S. homebuilders, is integrating XFRA into new construction from the start. SPAN's stated target is gigawatt-scale capacity by 2027. Whether that combination performs as promised at that scale is a question the next 18 months will begin to answer.

A product manager at a mid-sized SaaS company spent three hours with Claude Code trying to build an internal data pipeline. The AI did not fail. It produced a complete, working implementation along with a precise deployment sequence. She got through the first two steps, got confused by the credential format on the third, closed the terminal, and opened a ticket for engineering. The pipeline sat unbuilt for two weeks.

This is the failure mode that the prompting narrative has largely missed. The popular story about AI productivity stops too early in the workflow. Getting good output from an AI agent requires skill. But prompting is only half the loop. Once the AI returns a response, a human being has to execute the steps. That is where most of the failure occurs.

Cognitive scientists have spent decades studying why people fail to complete multi-step instructions. Working memory is the primary constraint. A 2020 review in the American Journal of Pharmaceutical Education found that working memory capacity is the central variable in instruction-following ability, and that people with lower working memory are particularly prone to starting sequences and failing to complete them. The breakdown usually happens somewhere in the middle, not at the start.

The modality of instruction matters too. Reading a sequence of steps in a chat window is close to the worst possible format for retention. It is linear, it is passive, and once you close the window to open your terminal, the instructions are gone. A 2016 study in Memory and Cognition found that physically performing each step at the moment of instruction dramatically improved both retention and completion rates. The habit of reading ahead undermines the very completion it is meant to prepare for.

The metric that matters is completion: what percentage of AI-generated instruction sequences actually get executed end to end by the person who received them? The answer, in most organizations, is lower than assumed. The gap between "AI generated a plan" and "the plan was implemented" is where the return on AI investment disappears.

The profile that survives is not the best programmer. It is the person who can read a ten-step deployment sequence, hold it in working memory, execute each step sequentially without losing track, recognize when a step has produced an unexpected result, and adapt without abandoning the task. Organizations that figure this out will stop asking "who can prompt well" and start asking "who can follow through." Those are related questions but they are not the same question.

The large enterprise AI story has a familiar shape. A global bank or a Fortune 100 manufacturer signs a multi-year deal with a consulting firm, launches a transformation program, forms a steering committee, and eighteen months later the AI tools are in a pilot with 300 users. It is how very large organizations manage risk with very large investments. But it leaves a wide swath of the economy on the sidelines.

Anthropic's new venture is built around the observation that mid-sized companies have compelling AI use cases, lack in-house engineering capacity to pursue them, and are not well served by the existing delivery ecosystem. The structure of the deal makes the bet explicit: the new firm is backed by private equity and alternative asset managers whose portfolios are full of exactly these kinds of companies.

The founding capital comes from Blackstone, Hellman and Friedman, and Goldman Sachs, each reportedly committing $300 million alongside Anthropic's own contribution. The broader consortium adds Apollo Global Management, General Atlantic, Leonard Green, GIC, and Sequoia Capital. That investor list is not incidental. Each firm has a portfolio of operating companies. Apollo alone manages assets across hundreds of businesses in industrials, financial services, and healthcare. The new firm has a built-in client pipeline from day one.

The deployment model mirrors what Palantir built with its Forward Deployed Engineering approach: engineers embedded directly with customers, building around the specific operations and workflows of each organization, rather than selling a standardized product and leaving implementation to the customer. A typical engagement starts with a small team sitting with the customer to identify where Claude can have the most impact, then building tools around that knowledge.

Mid-sized enterprises make decisions faster, have shorter internal approval cycles, and when leadership decides to adopt a technology, adoption actually happens on a timeline measured in weeks rather than years. There are far more mid-sized companies than large enterprises, they are distributed across every sector, and they are ready to move. A community bank that adopts Claude for loan documentation and compliance review is not a smaller version of JPMorgan's AI transformation. It is a faster, more direct path from contract to production deployment.