Every AI job starts with a number nobody can defend. How many GPUs, how much memory, how long the walltime. Ask for too little and the job dies hours in. Ask for too much and expensive hardware sits reserved and idle while the next job waits in the queue. @ismaeel_bashir_ lived that ritual from the inside. In his master's year at Edinburgh's parallel computing centre, he built a model that predicted what a job on the national supercomputer would need before it ran. Then he went to run machine learning at one of the world's largest quant hedge funds, and the same guessing game was waiting there, except the stakes were GPUs nobody can buy. In October 2025 he founded @ExpanseCompute with three fellow Edinburgh graduates, @nkdem_b, @erenzmendi03 and @yafet_melake. Two of them were still students. On September 16, Expanse announced a $5.3M seed round led by @crane_vc, with PXN Ventures and angels including former DeepMind researchers and AI infrastructure leaders, per HPCwire. The thesis fits in one of Bashir's lines: "The machine should carry the uncertainty, not the person." Expanse predicts exactly what a workload needs before a single GPU is committed, reading the job's code, the cluster's record of similar runs and the available hardware. It runs inside the customer's own environment, on Slurm, Kubernetes and the major clouds, so code and data never leave. The number that stuck with me is from the company's own launch post. In one cluster, Expanse measured nearly $8M of idle compute capacity in a single month. Not a projection. Measured. The backdrop makes the timing obvious. Industry estimates put roughly 30% of cloud spend lost to over-allocation, and Microsoft Research has reported GPU utilization around 50% across many internal deep learning workloads. GPUs are the scarcest resource in the industry, and half of them are being paid to wait. Crane wrote this check on September 16, then showed up in Outerlimit's $16M pre-seed a week later. Two infrastructure seed bets in six days tells me Crane has decided where the next margin in AI compute gets made. Congrats Ismaeel, Nikodem, Eren and Yafet, and the entire Expanse team. Up and to the right my friends! 💥🚀💪

Most of those platforms are running Kubernetes underneath. You are paying a company to have the skills you are trying to sell.

90% of distributed systems interviews in 2026 come down to these 7 points: 1) Know your guarantees At-least-once vs at-most-once vs exactly-once, and what duplicates mean for billing, emails, and counters. 2) Idempotency everywhere Request IDs, idempotency keys, de-dupe tables, and why retries without this create phantom writes. 3) Partitioning and hotspots Pick a shard key, explain rebalancing, and show how one celebrity user can melt a single partition. 4) Consistency model tradeoffs Strong vs eventual, read-after-write, quorum reads/writes, and what clients observe during failover. 5) Failure handling Timeouts, retries with backoff+jitter, circuit breakers, and why infinite retries are a DDoS you wrote yourself. 6) Ordering and time Per-key ordering vs global ordering, out-of-order delivery, clocks vs sequence numbers, and how to handle late events. 7) Backpressure and overload Queue growth, consumer lag, load shedding, and making p99 stable by dropping work intentionally (not randomly).

Retries alone don’t make a call reliable. My usual default is: → One overall deadline around the entire operation → A timeout for each attempt → One initial attempt plus no more than three retries → Backoff and jitter between retries → A circuit breaker close to the dependency, where it can observe every call But the most important rule should come before all of them: I only retry operations I know are safe to repeat. A timed-out request may still have completed downstream. Retrying it without idempotency could charge a customer twice, create two orders, or repeat another side effect. Retries should absorb temporary failures, not create new ones.

A retry without a timeout isn't resilience, it's a thread leak with good intentions.

Introducing Prime Sandboxes: MicroVM sandboxes purpose-built for RL training. Model training requires running tens of thousands of concurrent sandboxes, leading to complex and costly configuration. We built Prime Sandboxes for our own team. Today we're releasing them publicly.

Orchestrating agents at scale is a scheduling problem first, an intelligence problem a distant second.

A 309B-parameter model running at 49–59 tok/s. Not in a centralized datacenter. On a consumer Mac. Every one of these machines is a potential provider of open intelligence, not just a consumer of it. That's the part we're building toward. Rapid-MLX 0.15.0 👇

From live events and filmmaking to immersive experiences and Gaussian splats, @rendernetwork is powering new ways for artists and creators to bring their work to life. This Sat, @SpenserFX takes the stage at @Mo_Plus_Design NYC to share more about how artists are using the network across these workflows. Get tickets: motion-plus-design.com/home

Training is a relative race. Performance is judged against whatever the leading labs are running, and interconnect and memory gate the biggest runs. Inference is an absolute threshold: Hit the latency target at an acceptable cost, or don’t.

What if your AI agent had its own computer? Claw runs inside OptimAI Core Node. Soon, with OptimAI Router, models can connect to that environment too. Agent → Claw → Router → Model → Core Node The agent gets a runtime. The node gets intelligence. The network gets useful compute. This is where our idea of Agentic DePIN starts becoming real. We're not just decentralizing compute. We're connecting distributed compute to agents that actually need to use it. optimai.network

AI COMPUTE IS BECOMING A FINANCIAL MARKET The AI race is moving beyond models. Underneath every model, agent and inference application sits the same physical bottleneck: AI compute. GPUs need power. Power needs data centers. Data centers need capital. As AI usage grows, that infrastructure is becoming a market of its own. The first shift is already happening. GPU capacity is being measured. Compute prices are being benchmarked. Hardware is being financed. Future capacity is being priced. CME is preparing H100 and B200 compute futures based on GPU rental benchmarks. ICE and NATIVX are developing energy-adjusted compute futures across workloads including training and inference. Ornn is building around compute pricing, spot markets and derivatives. The important change is simple: AI compute is moving from something companies simply buy into something the market can price and hedge. The financing layer is growing with it. NVIDIA has partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR on platforms designed to mobilize more than $500B in third-party capital for AI infrastructure over time. Crypto is building its own financial rails around the same infrastructure. ✧ @akashnet turns distributed GPU capacity into an open marketplace. ✧ @bittensor uses token incentives to coordinate specialized compute, inference and digital work. ✧ @USDai_Official brings stablecoin liquidity and GPU-backed lending into the infrastructure layer. ✧ Venice uses DIEM to turn recurring inference access into a transferable onchain asset. ✧ x402 enables software and AI agents to pay for digital services programmatically. These are not the same products. They are different pieces of an emerging AI compute economy. And compute has an important difference from most financial assets: unused capacity cannot be saved. If a GPU is available for one hour and nobody uses it, that hour is gone. It cannot be carried forward and sold as another GPU-hour. That makes utilization, pricing and financing central to the business. Hardware depreciates. Electricity costs move. New chips arrive. Demand shifts between training and inference. Capital therefore needs better ways to price future compute, finance hardware and manage the risk around changing costs. That creates a natural progression: GPU supply → compute pricing → credit → futures → collateral → tokenization → liquidity This is where crypto’s role becomes much larger than decentralized cloud infrastructure. DePIN can coordinate the supply. Stablecoins can move the capital. DeFi can finance the infrastructure. Tokenization can package future access. Derivatives can manage price risk. Onchain settlement can connect the market. The endgame is not simply putting GPUs onchain. It is making AI compute a programmable financial resource: measured → priced → financed → hedged → tokenized → traded → settled AI creates the demand. GPUs provide the capacity. Capital finances the infrastructure. And crypto is beginning to build the financial market around the compute powering AI.

The most overlooked shift in demand-driven networks isn't the data architecture. It's the profile of the node operator. Traditional DePIN often rewards hardware uptime. You deploy a device, connect it, and wait for emissions. @vangrid_io introduces a different supply-side behavior. When demand is expressed through USDC bounties tied to specific locations, simply keeping hardware online is not enough. The sensor has to reach where the demand exists. That changes the operator's job. Instead of optimizing only for uptime, operators start optimizing for location, distance, timing and availability. The network stops looking like a static server farm. It starts looking more like a decentralized logistics fleet. That's a significant behavioral shift. Passive yield requires capital. Active spatial capture requires capital plus coordination. And that creates the real economic test: Are localized spatial data bounties valuable enough to compensate operators for the additional operational friction? Because once physical movement becomes part of the supply side, the network isn't just coordinating data anymore. It's coordinating labor, hardware and location.

yeah, masternodes are the quiet workhorses, but rewards only matter if uptime and governance are real tbh

Quorum turns a hardware failure into a rounding error instead of an outage. That's the design win.

Incentive alignment is a consensus problem in disguise. The protocol is the coordinator.

Physical AI has a data problem: not all data needed to train a robot comes from the same place. There is manipulation data: joint trajectories, teleoperation recordings and robot demonstrations that teach machines how to interact with objects and perform tasks. Then there is spatial ground truth: the streets, loading docks, stores, homes, warehouses and hallways where those machines actually have to operate. The first problem can be attacked with controlled environments. Put a robot in a lab, give it a task, record the movements and repeat the process at scale. The second problem is much harder. No robotics lab can realistically send employees and rigs across every street, building and commercial space on the planet, then repeat those captures every time the physical environment changes. @vangrid_io approaches that missing piece by turning existing smartphones into distributed sensors that can capture the environments outside those controlled facilities. A contributor can capture a real location, with multiple viewpoints providing the raw information needed to reconstruct its spatial structure. That data can then become structured ground truth for systems that need to understand physical environments beyond the boundaries of a robotics lab. And the scale is fundamentally different. A warehouse has a fixed location, a limited number of rigs and a finite number of environments it can reproduce. A distributed network of smartphones can move through cities, buildings, roads and businesses that would be impractical for a centralized data-collection operation to cover. The manipulation layer teaches a robot how to act. Spatial ground truth helps it understand where it is acting, what surrounds it and what that environment actually looks like.