Canonical
Buying intent
64 tracked signals | Top 15 topics are below | Sales and Engineering are carrying most of it.
Attention by team
LinkedIn activity, by teamWhere Canonical's own people are actually spending their attention, by team, by topic. Bands run Low to High against the busiest pairing on this page, and each cell also shows how much of that team's own activity it represents.
Topics being researched
30-day windowEvery tracked topic, ranked by volume, not by our guess at what matters. Confidence is the classifier's own certainty that a signal belongs where we've filed it.
3d ago
14d ago
14d ago
12d ago
9d ago
12d ago
16d ago
12d ago
22d ago
16d ago
20d ago
26d ago
17d ago
17d ago
17d ago
Need intent for a specific topic or industry?
We track the full taxonomy across every account in the graph — including themes not shown on this page.
Who's active at Canonical
verified title on fileTitles, seniority and topic straight from each person's own activity, with a LinkedIn link so you can check any of them yourself.
See everyone, not just the first 10
19 people across every department at Canonical, plus a LinkedIn profile link for each.
Primary products / business lines
LinkedIn company profileCanonical provides secure and reliable open source solutions, specializing in commercial support for Ubuntu. They are the company behind the popular Linux OS, Ubuntu, and lead a global community of 200,000 contributors.
Top accounts researching Canonical
names withheld on the public pageThese are companies whose own people brought up Canonical unprompted, not accounts we guessed might be interested. We can't yet tell an implementation partner from a genuine buyer here, names unlock along with the buyer profile below.
158,882 companies · 951,536 people are researching Artificial Intelligence
Canonical's own team shows 9 signals on this topic. No one outside Canonical has been seen researching the company by name yet — so this is the market it sits in, not a list of its buyers.
- Open Source13,333 cos · 49,179 people
- Cloud Infrastructure3,067 cos · 11,425 people
Buyer profile
company size · seniorityCompany size and how senior the people involved are, the two things that decide whether this is a real deal. Competitor overlap isn't computed yet for this account.
What's been said
public posts by Canonical's teamNo public post naming Canonical has surfaced in the past year, so this is what Canonical's own team is posting about publicly — their topics, in their words.
The GPU roadmap is starting to look less like a chip roadmap and more like a power roadmap. A few years ago the conversation was about getting enough accelerators. Now it is whether the facility can actually deliver the power, cooling and rack density those accelerators demand. The trajectory is striking: Hopper racks drew around 40 kW. Blackwell pushed past 120. Rubin lands near 190 to 230, with 600 kW racks specified for 2027. By the end of the decade, 1 MW-class racks are no longer a theoretical edge case. At that point, power is not a utility sitting behind the compute. It is part of the compute architecture. The implications run deep: You need the right DC distribution. You need liquid cooling designed for far higher thermal density. You need rack infrastructure that can carry the current draw. You need a facility that evolves with the silicon roadmap instead of stranding one generation later. This is why the next phase of AI infrastructure competition may be decided less by who announces the most GPUs, and more by who can actually deploy them at scale. GPU availability wins attention. Power architecture decides whether the GPU can be deployed. So for anyone planning capacity now, the question is no longer "which generation are we buying?". It is "will the partner we choose to build with today support the generation after that?" #AIInfrastructure #DataCentres #SovereignAI
Aug 2026A GPU count tells you what you bought. It does not tell you what the system can actually sustain once real agents are running with long contexts, repeated tool calls, persistent state and multiple workflows in parallel. For that world, the scarce resource may not be the accelerator itself. It may be how many useful workflows the system can run concurrently at a defined service level. And that number depends on far more than GPU count: GPU memory, KV cache, model architecture, context length, quantisation, concurrency, latency target Change any one of those and the real capacity of the system changes, while the GPU count stays exactly the same. That is why I think we may need a different way to talk about agent infrastructure. Instead of: GPUs installed or even: tokens per second perhaps the more useful metric is: successful concurrent workflows at a defined SLA One is a supply metric. One is a speed metric. The last one is a business metric. That distinction matters because a buyer can map successful workflows to productivity, revenue or cost avoidance. It is much harder to map a raw GPU count to any of those. But I think the direction is worth debating. As agents become the workload, should we keep measuring infrastructure by the hardware installed, or by the amount of useful work it can sustain? #AIInfrastructure #AgenticAI #InferenceEconomics
Aug 2026Cheap inference will not commoditise all AI infrastructure. It is more likely to make premium infrastructure matter more. Here is the counterintuitive part. As the price per token falls, AI stops being a demo and moves into revenue-critical workflows: trading, claims, diagnostics and customer operations. Each unit becomes cheaper. In aggregate, the system becomes mission-critical. The moment a workflow becomes mission-critical, the questions change. Buyers stop asking only how cheap the token is and start asking: Predictable latency: real-time work cannot wait Reliability: the workflow must not stall Data sovereignty: who controls the data and under which law Availability: capacity when it is needed Model choice: the right model for each task Security: controls an auditor will accept Cost predictability: a bill the business can plan around None of these are about the sticker price of intelligence. They are about whether the organisation can depend on it. Cheaper inference therefore does not flatten the infrastructure market. It separates it: Commodity capacity for experimentation, and dependable, governed infrastructure for the production workloads a business actually runs on. For regulated industries, the bar is higher still. “It usually works” is not a position that can be defended to a regulator, an auditor or a customer. The more intelligence a company consumes, the less it can tolerate infrastructure it does not control or cannot rely on. If AI is moving into your critical workflows, which of these would you refuse to compromise on? #SovereignAI #AIInfrastructure #InferenceEconomics
Aug 2026A genuine full-circle moment. I spent a significant part of my career working closely with the semiconductor and electronic systems ecosystem. Today, at Era4, I am helping build the sovereign AI infrastructure that the UK’s next industrial era will run on. Those two worlds are now converging. What Cadence and Hewlett Packard Enterprise are enabling through digital twin matters because AI infrastructure can no longer be designed using static assumptions. Power density, cooling, airflow, network behaviour, rack layout and workload performance all interact. Errors become expensive once the hardware is installed and energised. Using Cadence Reality Digital Twin with the HPE AI Mod POD and NVIDIA accelerated computing, ERA4 can model, validate and optimise the complete environment before the first rack is powered on. The result is lower deployment risk, faster time to service and better tokens-per-watt performance. It is also a clear sign that AI infrastructure is becoming a full-stack engineering discipline. The advantage is not simply access to GPUs. It is the ability to design, deploy and operate the entire system with greater predictability. Thank you to Hewlett Packard Enterprise, Nick Hume and Antti L. Lautennen at Cadence for helping bring this together, as well as the wider Cadence and ERA4 teams involved. From semiconductor design to sovereign AI infrastructure, it is rewarding to see these capabilities come together in a real deployment. If you could design out one risk before committing an AI cluster, would it be power, cooling, networking, performance or deployment time? #SovereignAI #AIInfrastructure #DigitalTwin Link to the full Video and article from HPE Discover 2026: https://lnkd.in/eDBDU8Dy
Jul 2026The price of intelligence is collapsing and enterprise AI budgets are still rising. The second is caused by the first. This is the Jevons paradox, and we have seen versions of it throughout computing: when the cost of a unit falls, we rarely settle for doing the same work more cheaply. We find ways to do far more. Watch what happens as the price per token declines: Run more workloads: usage expands as more applications become economically viable. Build more complex systems: multi-step reasoning, tool use and multi-agent workflows become practical. Automate the previously uneconomic: tasks that could not justify the cost suddenly enter scope. Embed AI into every workflow: inference moves from an occasional feature to part of the operational fabric. The result is falling cost per unit alongside rising aggregate consumption. For infrastructure planning, cheaper inference is not a reason to expect demand to soften. It is the mechanism through which demand expands, pulling a long tail of previously uneconomic workloads into production. The binding constraint then moves. It begins with model cost, shifts towards available compute and, at sufficient scale, becomes a question of power, cooling and grid capacity. So a dashboard showing cost per token falling is good news, but it is not the whole story. The real question is whether consumption is growing faster than unit costs are falling. If intelligence becomes cheap enough to embed in every workflow, what happens to compute demand, and are we planning for the scenario in which it rises rather than falls? #AIInfrastructure #InferenceEconomics #EnterpriseAI
Jul 2026There is no universally best GPU. There is only the right GPU for the workload you are actually running. Too many infrastructure decisions still default to the newest part. It feels safe, but it can also be the most expensive way to buy compute. Each NVIDIA architecture addressed a different set of bottlenecks rather than simply being “better” than the one before it. Volta introduced Tensor Cores and mixed precision. Turing targeted low-precision inference. Ampere expanded the training envelope. Hopper added FP8 and the Transformer Engine. Blackwell introduced NVFP4. Vera Rubin is a rack-scale co-design, not simply a faster standalone chip. That history matters commercially because infrastructure fit is workload-specific. A model processing billions of lightweight embedding requests may not need a flagship accelerator. Put it there and you risk paying premium rates for commodity work. A long-context reasoning workload places much greater pressure on memory capacity, bandwidth and KV-cache management. In that environment, memory architecture can matter as much as the headline compute number. Those are different purchases. The useful question is therefore not which GPU is newest. It is whether the system clears seven tests: Memory fit. Latency. Throughput. Interconnect. Power. Software maturity. Delivered cost per output. The last one reaches the P&L: cost per token, per inference, per completed job or per successful workflow. Alongside it sits the operational efficiency measure: tokens per watt. I spent years close to silicon before moving into cloud and AI infrastructure, and the pattern still holds: The spreadsheet does not care which generation you bought. It cares about utilisation, throughput and watts. No financial model has ever had a row for newest. Swipe for the architecture timeline and workload map. If you are buying AI infrastructure this year, are you specifying by generation or by workload? #AIInfrastructure #Inference #GPU
Jul 2026The AI infrastructure question is shifting from "How many GPUs?" to "What is the right compute for this workload?" Most infrastructure conversations assume every AI workload needs the largest accelerator available. The data says otherwise. Look at what developers actually run. Of the ten most-downloaded open models, only one is a text generator. The rest are compact embedding, ranking, classification and speech models. Their economics are set by memory bandwidth, latency, concurrency and data movement, not peak training throughput. That breaks the instinct to put everything on premium GPUs. A model handling billions of lightweight embedding requests belongs on a CPU or a cost-efficient GPU. Put it on a flagship accelerator and you pay premium rates for commodity work. So the advantage moves beyond securing GPU capacity. It comes from matching each workload to the right infrastructure: CPUs for highly concurrent, lightweight tasks Cost-efficient GPUs for scalable inference Premium GPUs for demanding models Memory-rich nodes for large-model serving Storage-optimised clusters for data-intensive pipelines Edge infrastructure where latency and locality matter The hyperscalers already run this way. Microsoft lifted inference throughput on its most-used models by 40 percent and improved tokens per dollar by more than 30 percent, largely by matching hardware to workload. This is a full-stack optimisation problem. Orchestration, workload placement, networking, storage and inference economics now matter as much as raw GPU inventory. The winners will not be those with the most powerful chips. They will be those that consistently run each workload on the right infrastructure, at the right cost. Are you buying the most powerful chip, or the right chip for each workload? #AIInfrastructure #Inference #Gpus
Jul 2026The cost of running an AI model can fall by up to 100x. Not one of the four biggest reasons is a smarter model. We spent three years asking whether models would keep getting better. The sharper question in 2026 is whether inference gets cheaper. A single model can be queried billions of times a day, and every query carries a unit cost. I came up through silicon before I moved to cloud, so this is the layer I struggle to look away from. The four levers doing the heavy lifting, roughly in order of impact: Model optimisation, quantise and prune: 85 to 95 percent lower cost per token. Pure software, and the strongest near-term lever. 3D advanced packaging, hybrid bonding: 80 to 90 percent. Memory moved physically closer to the compute. Custom silicon, inference ASICs: 70 to 80 percent. Hardware shaped to the model, not the other way around. Co-packaged optics: the long-term wildcard, for when copper runs out of room to move data. Deploy them together over time and the reductions compound toward two orders of magnitude. Two numbers now run the economics. Cost per token decides whether an application is commercially viable. Energy per token decides whether it can scale at all. Most of these levers improve both at once. The strategic shift: as models commoditise, advantage moves down the stack to inference economics, and to where the tokens are physically produced. Cheaper tokens do not just protect margin. They decide which AI applications exist, and who is allowed to run them. If you run AI in production: which of these four are you pulling first, and what is blocking the other three? #AIInfrastructure #Inference #Semiconductors
Jul 2026Next week, 9,000 AI leaders land in Paris for RAISE. Most conversations will be about models. The one that matters is about compute. The bottleneck on enterprise AI has quietly moved. It is no longer the model. It is whether you can get power, GPUs, and a compliant place to run them this year, rather than in three. That is what I will be at the Carrousel du Louvre to talk about. If you are building AI that has to scale under real regulatory and cost pressure, those are the constraints worth comparing notes on. I will be there both days, 8 and 9 July. The people I want to meet are the ones solving this from the other side: founders scaling inference, enterprise infrastructure leaders, and AI-native organisations. If you are attending, let's connect in Paris. #RAISESummit #SovereignAI #AIInfrastructure
Jun 2026