Samsung and AMD demonstrated CXL memory pooling at the 2026 OCP Global Summit, letting AI servers borrow idle capacity from a shared pool. Pooling would move memory sizing from each server to the pool, so orders would follow combined cluster load.
Perspective Coverage
3 publishers
- Builder
- Builder 40%
- Operator
- Operator 32%
- Investor
- Investor 28%
Reality
- Evidence35
- Adoption
- Not scored
- Hype gap+20
- Incentives70
- Confidence65
HPE has launched ProLiant Gen13 servers for AI inference on sixth-generation AMD Epyc chips, with the eight-GPU DL585a model due in early 2027. Only the single-socket DL525 is due in 2026, so a Gen13 AI plan for this year has to start with it and add GPUs later.
Reality
- Evidence45
- Adoption
- Not scored
- Hype gap+25
- Incentives75
- Confidence65
Nvidia and Sapphire Ventures joined Reactor's Series A, lifting the world-model startup's funding to $74 million from $59 million. Nvidia's stake sits in the layer that runs the models, days after AMD agreed to buy model maker World Labs for $8.2 billion in stock.
Reality
- Evidence55
- Adoption20
- Hype gap+25
- Incentives60
- Confidence55
Clockwork raises $31 million for software that keeps GPU jobs running through hardware failures
Clockwork Systems raised $31 million and launched TorchSnap, a tool that snapshots distributed AI jobs so they can resume after hardware failures. The customer deployments it cites predate the new feature, so its value depends on how long a buyer's own restarts take.
Reality
- Evidence40
- Adoption35
- Hype gap+30
- Incentives70
- Confidence45
Nebius bought Inferize, a 10-month-old Israeli startup with 17 staff, for an estimated $100 million to $150 million, according to Calcalist. The money goes on software that keeps inference GPUs from sitting idle, and Nebius is still adding data center capacity alongside it.
Perspective Coverage
3 publishers
- Builder
- Builder 38%
- Operator
- Operator 32%
- Investor
- Investor 30%
Reality
- Evidence55
- Adoption12
- Hype gap+30
- Incentives70
- Confidence60
buildOne report1 publisher Gimlet plans 100 MW of Cerebras wafer-scale capacity to split inference by phase
Gimlet Labs plans 100 MW of Cerebras-powered inference capacity, pairing wafer-scale chips with GPUs for different phases of each request. For buyers, the quoted 3,000 tokens per second is a company target until measured results appear.
Reality
- Evidence40
- Adoption15
- Hype gap+40
- Incentives70
- Confidence40
buildOne report1 publisher Nvidia's MaxLPS trial at Nscale ran 192 GPUs against a 140-GPU static baseline
Nvidia says DSX MaxLPS power sharing fits up to 40% more GPUs in the same approved power budget, citing an Nscale trial of 192 GPUs against 140. The gain holds only while GPUs rarely peak together, and each operator has to measure that on its own fleet before buying hardware against it.
Reality
- Evidence45
- Adoption12
- Hype gap+25
- Incentives80
- Confidence50
DeepSeek weighs 160,000 Huawei Ascend chips for a 1GW inference site in Ulanqab
CEO Liang Wenfeng said in investor discussions that Huawei starts delivering AI training chips around Q4 2026 or Q1 2027, and the Inner Mongolia fleet under consideration is 160 times the Ascend cluster DeepSeek ran in June 2026.
Reality
- Evidence27
- Adoption20
- Hype gap+44
- Incentives58
- Confidence34
buildOne report1 publisher China Merchants Bank's 60% token-cost cut outpaces the utilization gain it credits
The bank pooled nearly 10,000 heterogeneous accelerator cards under Kubernetes and reports average utilization rising from 35% to more than 60%. That density supplies 1.71 of the 2.5x drop in cost per million tokens.
Reality
- Evidence40
- Adoption55
- Hype gap+25
- Incentives65
- Confidence45
Huawei's 2035 token forecast assigns more than 90% of the traffic to agents
The company published the number days before its Connect 2026 show, alongside ten directions that name the cluster hardware, memory systems and chip method it sells. The agent share is the part a budget has to absorb.
Reality
- Evidence38
- Adoption24
- Hype gap+48
- Incentives86
- Confidence57
buildOne report1 publisher Open-weight safety pact hints Baseten could test controls in its runtime, but no division of labor is confirmed
Base Labs has signed Hugging Face and Goodfire to publish methods for training, evaluating and monitoring open models, with the check running where the activations are. The partners have not published an evaluation suite yet.
Reality
- Evidence42
- Adoption10
- Hype gap+38
- Incentives72
- Confidence48
Delos Data builds its inference server around an OAM socket buyers fill themselves
The startup has closed more than $100 million three months after debuting its first appliance, a four-accelerator box whose 36 OSFP ports are reported to give it network capacity comparable to Nvidia's Rubin systems.
Reality
- Evidence28
- Adoption12
- Hype gap+45
- Incentives70
- Confidence45
buildConfirmed2 publishers Delos Data's Nonstop AI Reference Architecture is meant to hold a mixed domain of GPUs, CPUs, memory and storage together through hardware and link failures. The 10x figures attached to it describe silicon the company still calls planned.
Reality
- Evidence42
- Adoption15
- Hype gap+34
- Incentives82
- Confidence55
buildOne report1 publisher Lambda ran 19 Blackwell nodes on the power budget provisioned for 16
The one measured number behind NVIDIA's tokens-per-megawatt pitch comes from Lambda's Blackwell cluster, where power reclaimed from static provisioning ran three extra nodes. Amazon's Annapurna Labs and d-Matrix got a line each.
Reality
- Evidence33
- Adoption45
- Hype gap+44
- Incentives90
- Confidence60
Xunce's parent guarantees every yuan of a 10 billion yuan compute loan
Shenzhen Xunce told the Hong Kong exchange that a subsidiary will seek up to 10 billion yuan of syndicated loans for an AI inference centre, with the parent guaranteeing the same amount because the project company cannot borrow on its own credit.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+22
- Incentives50
- Confidence52
Arm now sells a chip, and its licensees are buying cores from a competitor
The AGI CPU is a 136-core Neoverse part co-designed with Meta. Arm's pitch to custom-silicon customers used to be that it had nothing of its own to sell.
Reality
- Evidence40
- Adoption22
- Hype gap+32
- Incentives72
- Confidence45
I Squared's Saragon buys 53MW of ex-Sprint switch sites and calls it an AI inference platform
A $1bn platform seeded with ten converted Cogent facilities takes about a quarter of Cogent's installed power from ten of its 185 sites. The bet is on interconnection, not slab.
Reality
- Evidence45
- Adoption30
- Hype gap+45
- Incentives70
- Confidence50
AMD's Korea test lab is a bet that inference stops being a GPU-only purchase
AMD will open an AI Center of Excellence in Korea by year end to co-test its CPUs and GPUs with domestic NPUs. What is committed so far is engineers and a steering committee, not capital.
Reality
- Evidence50
- Adoption12
- Hype gap+18
- Incentives65
- Confidence45
Inference cost is now a storage and power problem, and it prices differently per workload
A vendor panel at Supermicro's Open Storage Summit argues GPUs no longer set token cost alone. The useful corollary: chat, batch and agentic work each need their own capacity plan.
Reality
- Evidence28
- Adoption18
- Hype gap+34
- Incentives86
- Confidence36
The chipmaker is discounting with dilution rather than price: nearly 1.4 million shares vest in year one, the rest in tranches tied to every $500 million of chips Google orders.
Reality
- Evidence66
- Adoption28
- Hype gap+32
- Incentives74
- Confidence63