Preston Holmes

"Just give me GPUs." That was then. This is now: Full-stack AI infrastructure, from ground to cloud. ABI Research ranked Nscale the #1 Neocloud, naming us Overall Leader, Top Innovator, and Top Implementer. Because building frontier AI takes more than GPUs. It takes an entire system designed to run them. The icing on the cake! Read the report here: https://lnkd.in/e2d7KhBm

83

2 Comments

*** Beyond the Black Box: Why Process Models Could Be the Operating System for Agentic AI *** We’re rushing to put AI agents into enterprise processes. But there’s a problem. AI without governance scales chaos. LLMs excel at reasoning and adapting to unpredictable tasks. But when agents make decisions inside mission-critical processes, “let the AI figure it out” isn’t an architecture. That’s where structured process modeling becomes interesting. For years, process diagrams have been treated as a way to document workflows. In the agentic AI era, we need to look at them differently: The process model can become the execution contract between humans, systems, and AI agents. And that changes how we should design processes. An AI agent needs its own accountability lane If an AI agent is making decisions, give it a dedicated lane in your workflow. Not a generic “Automation” lane. You should be able to see clearly: What did the AI decide? What did the system execute? What did the human approve? That separation matters for auditability and accountability. Every agent needs a fail-safe AI can stall, fail, run late, or produce unusable responses. Don’t just drop an AI task into a workflow and hope for the best. Use timeouts and error handling to define what happens when the agent doesn’t behave as expected. If the AI fails, the process should know what to do next. Put humans behind the right decisions Not every AI decision should go straight to production. For high-risk or high-value decisions, introduce a validation gate. AI makes the recommendation. The process evaluates it. If validation is required, a human takes over. Human-in-the-loop shouldn’t be an afterthought. It should be part of the process design. Let AI reason dynamically—but inside boundaries AI reasoning isn’t always linear. It may need to perform several activities, repeat one, skip another, or determine the order dynamically. This is where flexible subprocesses become particularly useful. You can give the agent room to reason dynamically without giving it control of the entire business process. Freedom inside the boundary. Governance outside it. One more important point: A process diagram isn’t just a picture. Underneath the visuals is a machine-readable process definition. Ambiguity in the model becomes ambiguity in execution. If a human can’t determine the next step without clarification, why would we expect an AI agent to? Treat the process model like compiled code. Loose logic can create real execution failures. The future of agentic automation isn’t: AI agents everywhere. It’s: AI agents operating inside well-designed processes. Structured modeling gives us a language to define those boundaries. The goal isn’t to restrict AI. It’s to give AI enough structure to operate safely at enterprise scale. #AgenticAI #AIAgents #Hyperautomation #ProcessOrchestration #EnterpriseArchitecture #EnterpriseAI #WorkflowAutomation

3

BREAKING: Ramp just published its list of top software vendors for February 2026 (Spend data, no opinions) 1️⃣ 𝐓𝐫𝐞𝐧𝐝𝐢𝐧𝐠 𝐯𝐞𝐧𝐝𝐨𝐫𝐬 (breakout growth vs size) AI infrastructure dominates. Cerebras, Runware, Modal, Crusoe, Novita. This is teams locking in the plumbing early. 2️⃣ 𝐅𝐚𝐬𝐭𝐞𝐬𝐭 𝐠𝐫𝐨𝐰𝐢𝐧𝐠 𝐯𝐞𝐧𝐝𝐨𝐫𝐬 (new customers added) Mostly generative AI. Anthropic, Cursor, ElevenLabs, Replit, OpenAI. Tools that sit directly inside daily workflows.

10

1 Comment

We found that fine-tuning a language model doesn't just add new features — it reorganizes existing ones. Inside a Marcus Aurelius LoRA, the real adaptation lives in co-activation clusters of shared features, not individual interpretable units. Clusters produce causal effects 10× beyond any single feature. Link in comments.

1

1 Comment

My colleague Anand has been explaining difficult IT concepts in an easy to digest way for a very long time. Check out his explanation of this weeks new hotness: Jev. I try not to hype most technologies because it just creates a lot of Chicken Little scenarios…but I’m telling you it’s really important that you take a minute and learn about Jev. As a Type 1 model it can potentially save you an exorbitant amount of money as it often needs a fraction of the tokens of Type 2 models. If you have no idea what Im yammering on about, watch Anand’s video

6

The most recent Queued Up report from Berkeley Lab is out. Steven Zhang and I are proud to be listed as coauthors on this edition. The Interconnection.fyi team worked closely with Joseph Rand and the rest of the LBNL team to provide the data that powers this year's analysis. Supporting this level of research is core to our mission of bringing transparency to the wholesale energy markets. The 2025 edition (covering data through the end of 2024) highlights some significant shifts in the landscape: Active Capacity: 2024 closed with nearly 2,300 GW of generation and storage seeking interconnection. Changing Mix: Active natural gas capacity increased by 72% year over year, while solar and storage saw slight decreases in total queue volume. The Backlog: 408 GW of capacity already has an executed or draft interconnection agreement but has not yet reached commercial operations. Timelines: For projects built between 2018 and 2024, the median duration from request to operation has doubled compared to the early 2000s. This report is the definitive annual benchmark for the industry and provides a vital baseline as we begin to see the implementation of FERC Order 2023. The analysis in this report is based on our EOY 2024 data snapshot. Since then, we have continued to track and update these queues every single day. If you want a live view of how these numbers have shifted in the months since this snapshot was taken, follow Interconnection.fyi and subscribe to our Substack. You can find the full slide deck, interactive maps, and raw data files at the link in the comments.

42

3 Comments

When most people hear "solid-state," they think storage. But in Santa Clara, xMEMS Labs just gave that term new resonance, literally. The company just raised $21M in Series D funding led by Boardman Bay Capital Management, with Cloudview Capital, CDIB-TEN Capital, Harbinger Ventures, Susquehanna Asia Venture Capital, and other strategic investors tuning in. This isn't another round of capital, it's an #amplifier for a company turning silicon into a new language of #sound and #cooling for the AI era. Founded in 2018 by Joseph Jiang and Jemm Liang, xMEMS spun out of piezoelectric R&D at TSMC and turned what was once wafer-level theory into commercial hardware. Joseph Jiang, a veteran of Knowles, Fortemedia, and InvenSense, leads with precision and swagger, while Jemm Liang, who built Ultrachip from the ground up, drives the engineering with surgical focus. Co-Founder & CLO Wei-Fu Hsu, the former MediaTek exec and IP expert, ensures xMEMS' innovation is as protected as it is disruptive. Their play? A monolithic MEMS platform that transforms voltage into vibration, silicon into sound. Unlike legacy coil speakers and fans, xMEMS tech doesn't move parts, it moves markets. Over 500K MEMS speakers shipped in 1H'24, backed by >250 global patents. The Cypress & Sycamore lines redefine what "small" can do: 1mm-thin, full-range, and capable of sub-bass response once thought impossible in earbuds. Their XMC-2400 µCooling chip, 1mm-thin, fan-on-a-chip, inaudible, moves 39cc/sec of air at just 30mW. That's not cooling, that's choreography at a molecular level. Dr. Chester Hwang, named CTO on Oct 15, 2025, joins to push that curve further. Alongside veterans Mike Housholder (VP Marketing & Biz Dev), Jim Wargnier (VP N.A. & EU Sales), James Lee (VP & Korea GM), Steven P. Bentley (VP Sales), and Martin Lim (VP MEMS Tech), xMEMS is scaling at semiconductor speed. They're already in BleeqUp's AI sports glasses and teamed with Dongguan Rayking Electronics for TWS modules. When OEMs talk "next-gen," they're talking xMEMS-level integration. For investors like Boardman Bay's Will Graves, this isn't just sound engineering, it's signal intelligence. With AI pushing devices thinner, faster, and hotter, xMEMS' piezoMEMS platform sits at the intersection of #acoustics and #thermals, where every millimeter matters. Backing them is a bet that the next wave of #sensorytech won't just compute, it'll feel alive. So what's next? Scaling production, expanding capacity, and embedding xMEMS' silicon pulse into everything from earbuds to AI #datacenters. In a world chasing speed, silence, and fidelity, xMEMS isn't following the beat, they're building the frequency the future will hum to. #Startups #StartupFunding #VentureCapital #SeriesD #AI #Sound #SoundEngineering #Semiconductors #Data #DataDriven #Technology #Innovation #TechEcosystem #StartupEcosystem #Hiring #TechHiring If software engineering peace of mind is what you crave, Vention is your zen.

16

1 Comment

NVIDIA's DGX Spark started shipping this week. $3,999 for a petaflop on your desk. Research labs currently skip experiments because justifying cloud spend takes three approvals. Startups burn runway on compute before knowing if ideas work. Teams prototype in the cloud, then find their models behave differently in production. Run 200B parameter models locally. Fine-tune 70B models without cloud dependency. Link two units for 405B parameters. The development cycle tightens considerably. The GB10 Grace Blackwell chip is purpose-built for this—MediaTek collaborated on the CPU side, 128GB of memory sits between the CPU and GPU without the usual bottleneck. Different from retrofitting gaming hardware for AI workloads. Training frontier models still requires massive infrastructure. Production inference at scale needs proper compute. But prototyping, testing, fine-tuning—that friction just dropped significantly. Useful ideas die when the barrier to experiment is too high. The path from concept to test matters. $4,000 isn't nothing, but it's orders of magnitude less than building a cluster. NVIDIA's framing this as "democratizing AI"—partially marketing, partially true. The number of teams who can now iterate at this level just expanded considerably.

77

4 Comments