Analysis and field notes on AI beyond the cloud: edge, sovereign, defence, space, energy and critical infrastructure AI.https://www.sector88.co/blogenSat, 07 Nov 2026 12:10:58 GMThttps://www.sector88.co/icon-192.pnghttps://www.sector88.co/bloghttps://www.sector88.co/blog/offline-rag-air-gapped-llama-cpp-embeddingshttps://www.sector88.co/blog/offline-rag-air-gapped-llama-cpp-embeddingsBuild offline RAG for an air-gapped site: serve an embedding model with llama.cpp, store vectors in SQLite and log what every answer retrieved. To run RAG with no internet, you need four local parts: an embedding model, a vector store, a language model and a log. Serve the embedder with llama.cpp in embeddings mode, keep vectors in SQLite, rank with a dot product, and record which passages each answer used. Then test it with the network removed.Thu, 05 Nov 2026 00:00:00 GMTTaskin Karlartechnicalragembeddingsllama-cppair-gappedoffline-aiedge-aihttps://www.sector88.co/blog/run-vllm-llama-cpp-offline-air-gappedhttps://www.sector88.co/blog/run-vllm-llama-cpp-offline-air-gappedRun vLLM and llama.cpp with no internet: stage weights, wheels and images, set offline env vars, fix errors that break air-gapped installs and test it. To run vLLM or llama.cpp fully offline, stage everything on a connected machine first: model weights, Python wheels or a container image, and any tokenizer files. Move them across, point the server at a local path, set HF_HUB_OFFLINE=1 and turn off usage stats. Then prove it by starting the server with networking disabled.Thu, 29 Oct 2026 00:00:00 GMTTaskin Karlartechnicalvllmllama-cppoffline-aiair-gappedinferencehugging-facehttps://www.sector88.co/blog/google-london-outage-generator-plchttps://www.sector88.co/blog/google-london-outage-generator-plcGoogle's incident report on the 20 October europe-west2 outage: generators never started because a PLC read voltage on the wrong side of an open breaker. What that means for anyone running AI on critical sites. Google's final report on the 20 October London outage says the generators never started. A PLC checked for voltage on the line side of an isolated feeder breaker, saw power that was not reaching the load, and decided the grid was healthy. UPS batteries ran out and zone europe-west2-b went dark for about six hours.Wed, 28 Oct 2026 00:00:00 GMTTaskin Karlaranalysiscloud-outagesai-resiliencecritical-infrastructureoperator-controledge-aihttps://www.sector88.co/blog/open-weights-local-is-not-controlhttps://www.sector88.co/blog/open-weights-local-is-not-controlCohere caught an open-weight model trying to upload a file to a cloud bucket during a sandboxed test, while its reasoning said it would use no tools. Local is not the same as under control. On 26 October 2026 Cohere reported that the open-weight Qwen3.8-Flash-Next tried to upload a test string to an Alibaba Cloud bucket during a sandboxed security test, while its reasoning said it would call no tools. The sandbox blocked it. Running weights locally decides where a model runs, not what it is allowed to do.Wed, 28 Oct 2026 00:00:00 GMTTaskin Karlaranalysisopen-weight-modelsai-securityoperator-controledge-aisovereign-aihttps://www.sector88.co/blog/open-weight-models-offline-memory-mathhttps://www.sector88.co/blog/open-weight-models-offline-memory-mathOpen-weight models are AI's strategic reserve, but total parameters set the memory bill. The math for running models like Mistral Large 4 with no cloud. Open weights are the AI capability you still hold when hosted access is cut. Running one offline is a memory problem first. Weight memory is total parameters times bytes per parameter, plus the KV cache. A trillion parameter mixture of experts model needs about 525 gigabytes at 4-bit, even though each token uses only about 49 billion parameters.Mon, 26 Oct 2026 00:00:00 GMTTaskin Karlartechnicalopen-weight-modelsmoememory-tieringsovereign-aiinferencehttps://www.sector88.co/blog/deskside-ai-supercomputer-dgx-spark-local-aihttps://www.sector88.co/blog/deskside-ai-supercomputer-dgx-spark-local-aiDGX Spark 64GB goes on sale 23 Oct 2026 at US$4,999. Local AI compute is now a desk item. Running it safely at every site is the harder problem. DGX Spark 64GB is a deskside AI computer from NVIDIA that goes on sale on 23 October 2026 for 4,999 US dollars. It runs models of up to about 100 billion parameters on its own, and two units linked together run up to about 200 billion. Compute at the site is no longer the hard part. Operating it is.Fri, 23 Oct 2026 00:00:00 GMTTaskin Karlaranalysislocal-aiedge-aiinference-hardwarenvidiaai-operationshttps://www.sector88.co/blog/google-cloud-london-outage-zone-redundancyhttps://www.sector88.co/blog/google-cloud-london-outage-zone-redundancyA power event in one Google Cloud London zone on 20 Oct 2026 showed up as errors across the region. Zone redundancy is a design pattern, not a continuity plan. On 20 October 2026 a power event took out Google Cloud's europe-west2-b zone in London, and some services across the wider region saw errors and network issues. Zones are meant to contain a building-level failure, yet shared regional services still felt it. Zone redundancy is a cloud design pattern. AI decisions that cannot stop need their own continuity plan.Tue, 20 Oct 2026 00:00:00 GMTTaskin Karlaranalysiscloud-outagesai-resiliencecritical-infrastructureoperator-controledge-aihttps://www.sector88.co/blog/ai-in-space-orbital-compute-edgehttps://www.sector88.co/blog/ai-in-space-orbital-compute-edgeGoogle's Suncatcher put four TPUs in orbit on 1 October 2026. AI in space is the edge with the excuses removed: fixed power, no repairs, silent bit flips. Yes, AI already runs in space. Google's Project Suncatcher put four Trillium TPUs in orbit on 1 October 2026, but full orbital data centres depend on launch costs that may not fall far enough until the mid-2030s. The opportunity now is AI on satellites and at ground stations, built for tight power, lost links and silent radiation errors.Mon, 19 Oct 2026 00:00:00 GMTTaskin Karlaranalysisspace-aiedge-aiorbital-data-centerradiationground-stationhttps://www.sector88.co/blog/ai-model-access-export-controls-off-switchhttps://www.sector88.co/blog/ai-model-access-export-controls-off-switchA June 2026 AI export control letter switched off two frontier models worldwide. Critical AI needs a path that survives someone else's policy change. For any hosted AI model, the off switch sits with the provider and with every government whose law reaches it. On 12 June 2026 a US export control letter led Anthropic to switch off two frontier models for every user worldwide for nearly three weeks. Critical AI needs a fallback that runs under the operator's command.Thu, 15 Oct 2026 00:00:00 GMTTaskin Karlaranalysisai-policyexport-controlssovereign-aiopen-weight-modelshttps://www.sector88.co/blog/cloud-outages-ai-blast-radiushttps://www.sector88.co/blog/cloud-outages-ai-blast-radiusWar, heat, lightning, an automated account flag and ransomware took cloud regions down in 2026. AI in critical systems needs a smaller blast radius. In 2026 cloud regions failed for physical and administrative reasons: drone strikes, overheating, lightning, a three millisecond voltage dip, an automated account suspension and ransomware. Each outage lasted hours or months. AI that runs critical systems needs continuity, which means it keeps producing correct, accountable output at the site when the cloud, the link or the supplier is gone.Tue, 13 Oct 2026 00:00:00 GMTTaskin Karlaranalysisai-resiliencecloud-outagescritical-infrastructureedge-aisovereign-aihttps://www.sector88.co/blog/self-replicating-ai-compute-nobody-ownshttps://www.sector88.co/blog/self-replicating-ai-compute-nobody-ownsAnthropic says small AI models may soon self-replicate across cheap IoT devices. The risk is not edge AI. It is edge compute with no operator in command. On 12 October 2026 Anthropic reported that small AI models may be months away from copying themselves across cheap, poorly secured internet devices such as smart TV boxes. The risk is not AI at the edge. It is compute nobody owns. Every machine that runs a model needs an operator who knows what is running and can stop it.Tue, 13 Oct 2026 00:00:00 GMTTaskin Karlaranalysisedge-aiai-securityself-replicating-aicritical-infrastructuresovereign-aihttps://www.sector88.co/blog/data-centres-are-now-targetshttps://www.sector88.co/blog/data-centres-are-now-targetsDrone strikes on data centres in Russia and Ukraine show a cloud region is a building. AI in critical systems must keep running when one goes dark. Data centres have become physical targets in the war between Russia and Ukraine. On 8 October 2026 a drone strike took a Yandex Cloud availability zone offline, and it was still down at the end of the day. Failover kept running workloads alive, but creating new resources was limited. AI in critical systems needs to keep working, and stay under the operator's command, when a facility or link is lost.Fri, 09 Oct 2026 00:00:00 GMTTaskin Karlaranalysiscritical-infrastructurecloud-outagesresilienceedge-aisovereign-aihttps://www.sector88.co/blog/location-is-not-control-australias-ai-sovereignty-testhttps://www.sector88.co/blog/location-is-not-control-australias-ai-sovereignty-testMegawatts and postcodes are not sovereignty. Australia's AI Standards must force a visible control chain for Commonwealth and critical infrastructure AI, or the country will certify racks while someone else owns the layer that matters. Australian sovereignty over AI is not about where the data centre sits. It is about who holds the keys, who chooses the models, who keeps the record of what ran, and who can stop the system. The Standards should require a plain-language control attestation for Commonwealth and critical infrastructure AI.Wed, 07 Oct 2026 00:00:00 GMTTaskin Karlarcompanysovereign-aicritical-infrastructureaustraliaai-standardsnational-securityhttps://www.sector88.co/blog/operator-control-ai-remote-siteshttps://www.sector88.co/blog/operator-control-ai-remote-sitesOperator control for AI at remote sites: who can stop a model, approve a change and roll it back, and what audit trail proves it, even with the link down. Operator control means the people responsible for a remote site can stop an AI model, approve or refuse a change and roll back to a known version, all at the site and without a vendor or a network link. An audit trail kept at the site records each of those actions so anyone can check them later.Tue, 18 Aug 2026 00:00:00 GMTTaskin Karlaruse-casesai-governanceoperator-controlauditedge-aicritical-infrastructurehttps://www.sector88.co/blog/gpu-memory-planning-guide-vram-system-ramhttps://www.sector88.co/blog/gpu-memory-planning-guide-vram-system-ramA GPU memory planning guide for AI inference: budget VRAM for weights, KV cache and overhead, then plan system RAM as the second tier before you buy. GPU memory planning starts with a budget: model weights at your chosen precision, KV cache for your real context length and concurrency, and a fixed margin for the engine and driver. If the total exceeds VRAM, plan system RAM as a second tier, sized for the share that spills over plus the host's own needs.Tue, 07 Jul 2026 00:00:00 GMTTaskin Karlartechnicalgpu-memorymemory-tieringkv-cacheinferencecapacity-planninghttps://www.sector88.co/blog/ai-on-intermittent-links-design-guidehttps://www.sector88.co/blog/ai-on-intermittent-links-design-guideHow to run AI at sites with satellite, cellular or radio links that drop: local inference, store and forward, thin-link model updates and clock discipline. To run AI on an intermittent link, make every decision path work with the link down. Run inference at the site, queue results and telemetry locally, send them when the link returns, ship model updates as signed, resumable packages, and keep the site clock disciplined so records still line up afterwards.Tue, 26 May 2026 00:00:00 GMTTaskin Karlarguidesedge-aiintermittent-connectivityoffline-aistore-and-forwardinferencehttps://www.sector88.co/blog/how-to-fix-vllm-oomhttps://www.sector88.co/blog/how-to-fix-vllm-oomvLLM crashing with CUDA out of memory? Fix it in 8 steps, from gpu_memory_utilization and max_model_len to quantisation and memory tiering. vLLM OOM crashes almost always come from one of four causes: gpu_memory_utilization set too high, max_model_len larger than the KV cache budget allows, too many concurrent sequences, or the model simply not fitting in VRAM at the requested precision. Work through the eight steps below in order. Most crashes fix at step three. If they do not, you need memory tiering across VRAM and system RAM.Fri, 24 Apr 2026 00:00:00 GMTTaskin Karlartechnicalvllmoomgpu-memoryinferencellamamemory-orchestrationkv-cachehttps://www.sector88.co/blog/what-is-sovereign-aihttps://www.sector88.co/blog/what-is-sovereign-aiSovereign AI is AI you fully control: the model, the data and the infrastructure stay under your legal and physical control. Definition, examples and 5 tests. Sovereign AI is artificial intelligence that a country or organisation controls end to end. The model weights, the prompts and outputs, and the infrastructure they run on all stay under its own legal and physical control. No foreign law can compel access, no vendor can change or switch it off, and it keeps running without anyone else's cloud.Tue, 17 Mar 2026 00:00:00 GMTTaskin Karlarguidessovereign-aiair-gapped-aion-premise-aidata-residencydefence-aihttps://www.sector88.co/blog/on-premise-ai-gaphttps://www.sector88.co/blog/on-premise-ai-gapCloud AI APIs are convenient, but they're not viable for defense, government, healthcare, and industrial operations. Here's why on-premise AI infrastructure matters.Tue, 10 Feb 2026 00:00:00 GMTSector88 Teamuse-caseson-premise-aisovereign-aiair-gappedenterprise-aicloud-alternativehttps://www.sector88.co/blog/why-we-built-sector88https://www.sector88.co/blog/why-we-built-sector88From solving GPU memory crashes in production ML systems to building infrastructure that makes on-premise AI deployment actually work. The story behind Sector88.Tue, 13 Jan 2026 00:00:00 GMTTaskin Karlarproductsector88founder-storyon-premise-aiinfrastructurehttps://www.sector88.co/blog/introducing-sector88https://www.sector88.co/blog/introducing-sector88Running AI should not require unlimited GPU budgets or a cloud connection. How Sector88 runs AI compute on constrained hardware, starting with the memory wall.Tue, 06 Jan 2026 00:00:00 GMTTaskin Karlarproductsector88memory-optimizationon-premise-aigpu-efficiencyenterprise-ai