# The Next Evolution of Kubernetes: GPU-Centric Infrastructure (Fujitsu) > [!question] Lingering Questions > - What is the project actually called? Auto-captions render it three inconsistent ways ("Codi," "Cording"), and none is searchable — find it via the CNCF Sandbox listing before citing anything. > - What hardware does composability actually require — commodity servers, CXL-capable platforms, or Fujitsu-specific hardware? Composable infrastructure has historically needed specific interconnects, and the talk never says. > - Zero measurements anywhere in a talk about resource optimization — what utilization improvement does composable infrastructure actually deliver? > [!tip] Article Ideas > - none flagged — a CNCF Sandbox announcement wrapped in an unevidenced cost argument; too little concrete material to build on until the project has a findable name and a number attached > [!link] Cross Reference > - [[AI Cost & Tokenomics MOC]] — naming electricity alongside GPU and memory as a primary infrastructure cost, third of three KubeCon Japan 2026 talks in this batch on physical AI-infrastructure constraints — cross-reference [[Infinite Agents, Finite Kubernetes (SoftBank)]] and [[How CNCF Projects United Two Countries' Engineers (NTT-SK Telecom)]] > [!quote] Vocabulary > - **Composable/disaggregated infrastructure** — hardware resources (GPU, memory) attached and detached dynamically to match workload, rather than statically provisioned per server ## Author's thesis AI platforms must shift from CPU-centric to GPU-centric infrastructure because of soaring GPU, memory, and electricity costs; the proposed mechanism is dynamic resource composition — software that scales composable hardware up and down to match workload. The news is that Fujitsu's composable-orchestration project has been accepted as a CNCF Sandbox project. The cost framing is specific and sound; the technical claims carry no measurements at all. ## Key arguments (≤3) 1. **The cost decomposition is specific and correct**: naming GPU, memory, *and* electricity separately as distinct cost drivers — treating power as a capital line item rather than an operating footnote — is what makes the efficiency conclusion follow rather than merely sound prudent. 2. **Dynamic composition is the right shape of answer**: static provisioning wastes the expensive resource during idle periods; composing hardware to workload attacks exactly that waste, at least in principle. 3. **The announcement is honestly scoped as early-stage**: CNCF Sandbox now, Incubating as the stated goal — not overclaimed as production-ready — but the technical claim underneath it (optimized utilization with efficient power consumption) has zero supporting measurement. ## Claim atoms ### Fujitsu's composable orchestrator enters CNCF Sandbox > [!quote] ID:rw_01kysr31wtpqefw94t11hyb7s1 > one of the biggest news is that it it has been approved as CNCF Sandbox project. **Where/why**: The announcement, roughly two minutes in. The project name is unrecoverable from auto-captions ("Codi"/"Cording") — find it via the CNCF Sandbox listing. Sandbox is the entry tier: foundation interest, not production validation. ## Permanent candidates - **Electricity belongs in the capital cost model, not the operating footnote** — once power is named alongside hardware as a primary cost of compute, efficiency stops being good practice and becomes the term that determines what the platform can afford to run. Strong candidate: cleanly generalizes and connects to the AI-cost/tokenomics thread already developing in the vault. ## Processing checklist - [ ] Look up the Fujitsu composable-infrastructure project in the actual CNCF Sandbox listing to get its real name and any published utilization data - [x] Connect into the relevant MOC — [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] ## Links - [[3. Resources/MOCs/AI Cost & Tokenomics MOC]] — electricity-as-capital-cost framing for AI infrastructure economics - [[3. Resources/Literature Notes/The Next Evolution of Kubernetes - GPU-Centric Infrastructure (Fujitsu)]] — source video (KubeCon + CloudNativeCon Japan 2026)