Keith Teare

🚀 Agentic AI Needs a New Infrastructure Playbook A new partnership between d-Matrix and Gimlet Labs, Inc. highlights a growing shift in AI infrastructure: heterogeneous compute. The companies plan to integrate d-Matrix Corsair inference accelerators alongside GPUs in Gimlet Cloud, creating a hybrid architecture designed for the next wave of real-time, agentic AI workloads. ⚡ The idea is simple: GPUs handle compute-heavy inference tasks Corsair accelerators handle memory- and latency-sensitive stages like token generation The result? Potentially up to 10x improvements in latency and throughput per watt compared with GPU-only deployments. 💬 Key perspectives from the announcement: “If you can fundamentally change how people interact with AI, they’ll be much more engaged.” — Zain Asgar, CEO, Gimlet Labs “Inference is never a one-size-fits-all problem. Heterogeneity is the path forward.” — Sid Sheth, CEO, d-Matrix “Platforms poised for success are those that developers can use without thinking about the hardware.” — Matthew Kimball, Moor Insights & Strategy 💡 Why this matters As AI moves from training models to production-scale inference, operators are running into two big constraints: 🔌 Power limits in data centers ⚙️ GPU inefficiency for memory-bound inference tasks Hybrid architectures that combine GPUs + specialized inference silicon may become the blueprint for scaling agentic, real-time AI systems—without blowing past power budgets. 📅 The joint platform is expected to reach select customers in the second half of 2026. 📊 Bottom line: The next phase of AI infrastructure may not be about bigger GPU clusters—but about smarter combinations of chips orchestrated by software. Full story at Data Center Knowledge: https://lnkd.in/eDu6Yf6f #AI #ArtificialIntelligence #AIInfrastructure #AgenticAI #Inference #DataCenters #Semiconductors #CloudComputing #GPUs #TechInnovation

44

2 Comments

🚀 Digital Twins + LLMs: The Next Breakthrough in Personalised Patient Care A new paper caught my eye today — and it highlights just how quickly AI is reshaping the future of healthcare. Researchers have developed DT-GPT, a large-language-model approach that can predict a patient’s health trajectory using their electronic health record — even when the data is incomplete, noisy, or sparse. The result? AI can now create digital twins that model a patient’s future health in extraordinary detail. 🔍 Why this matters Digital twins act as virtual patients, allowing clinicians to: ✅ Forecast disease progression earlier ✅ Test treatments safely in a simulated environment ✅ Understand how an individual might respond to an intervention ✅ Reduce adverse events before they happen ✅ Tailor care with precision that was impossible just a few years ago DT-GPT outperformed state-of-the-art models across lung cancer, ICU, and Alzheimer’s datasets — and even delivered zero-shot forecasting, meaning it could make accurate predictions without task-specific training. 🌟 What this unlocks This is a major step toward: Personalised treatment planning Virtual clinical trials AI-guided therapy selection Continuous patient monitoring Safer, more proactive healthcare The idea that we can “rehearse” treatments on a digital twin before treating the real patient is transformative. We’re moving from reactive care to predictive, model-driven medicine.

4

Aidoc just closed $150 million in funding to scale its clinical AI platform beyond radiology. Led by General Catalyst and Square Peg, with investment from Nvidia’s venture arm and four major U.S. health systems (Hartford HealthCare, Mercy, Sutter Health, WellSpan), this brings their total raised to $370M, including a $40M credit facility. Founded in 2016, Aidoc built its reputation on FDA-cleared radiology tools, pushing into intracranial hemorrhage, pulmonary embolism, and rib fractures. Now they’re expanding into cardiology, neurovascular, vascular, and even oncology use cases, all powered by their new CARE™ foundation model. It’s designed to build new clinical solutions up to 20x faster, with FDA-cleared examples already under its belt. #HealthTech #ClinicalAI #Startups

12

1 Comment

“Skills is all you need” 🚀 Now that we’re done with MCPs - let’s talk about skills… Check out Anthropic’s latest talk on agentic systems and Skills. It’s only 15 minutes and definitely worth your time: <https://lnkd.in/dKe2-Bci> It confirms a shift I’ve been seeing (and warning about) for a while. If you’re building in GenAI right now, the writing is on the wall: The frontier models are coming for your orchestration layer, and there isn’t much space left for defensible innovation. Here are my 3 key takeaways and what they mean for builders: **1. The agent layer is converging and being owned by frontier models** Companies spent the last year reinventing the same orchestration loops: planning, reflection, memory. This scaffolding is being commoditized and doesn’t require verticalization. The models scale nicely across use cases and industries. **2. “Skills” are all you need** 🧠 The “application layer” is turning out to be shockingly simple: text files. Anthropic’s Skills Framework proves it. A Skill isn’t a complex SaaS app - it’s often just a text file that encapsulates: 1. **Procedural Knowledge**: The specific steps to do a job 1. **Policies**: The guardrails of what not to do 1. **Tools**: The specific API connections required Whether this exact implementation wins doesn’t matter. The direction is clear. 💰 **My Bet**: Most enterprise skills are neither complex nor unique. If your startup is building generic “skills” - horizontal or vertical tasks that look the same across organizations - you’ll be commoditized by in-house teams or low-margin marketplaces. ⚠️ **3. The Steamroller is here** 🚜 OpenAI, Anthropic, and Google are determined to own the platform, not just provide the model. They’re systematically removing friction points others are rushing to solve: context compression, sub-agent management, planning flows. Too many startups confuse temporary friction (which models will solve) with inherent limitations (which need product solutions). That’s a pricey confusion. 🔬 **The Litmus Test for Founders**: Imagine your app as a simple “Skill” instead of a complex agent. Does it still hold value? If not, you’re building on the tracks while the train is coming. 🚂 **The takeaway**: Go to higher ground. Find a moat in proprietary data or true scientific/industry expertise. #GenerativeAI #ProductStrategy #VC #DeepTech #FutureOfWork

29

National AI Strategies Without Compute Strategies Are Wish Lists — A Frank Assessment Of Government Planning Around the world, governments are publishing AI strategies. Most contain familiar themes: Innovation. Research. Education. Skills development. Economic growth. Digital transformation. All important. But many overlook a critical question. Where will the compute come from? A national AI strategy without a compute strategy is like an energy strategy without power stations. Or a transportation strategy without roads. AI requires physical infrastructure. Every ambition ultimately depends upon: * Data centres * Energy generation * Semiconductor access * Connectivity * Cybersecurity * Sovereign storage * Skilled infrastructure teams Without these foundations, AI strategies risk becoming aspirational documents rather than operational realities. The challenge is not that governments lack vision. The challenge is that infrastructure planning often receives less attention than innovation planning. That imbalance is becoming increasingly difficult to sustain. The countries that succeed in AI over the next decade will likely be those that view compute as strategic infrastructure. Not merely as a technology service. This is where TerraVaultis MDCaaS provides a practical framework. Through sovereign-ready deployment, modular scalability, energy-efficient design, and rapid implementation, MDCaaS enables nations to translate AI ambition into AI capability. The future winners will not necessarily be those who write the best strategies. They will be those who build the infrastructure required to execute them. #TerraVaultis #MDCaaS #ModularDataCenters #AI #DataCenters #Infrastructure #PrivateEquity #SovereignAI #LiquidCooling

3