Prem Jonnalagadda

[Eridu @EriduAi]

AI scale-up infrastructure needs to move beyond the rack. Today's scale-up architectures are reaching their limits, requiring a new approach.

AI workloads are growing and evolving in ways that are straining the network--more model parameters and active experts, more data movement for communication collectives, less predictable traffic, and greater demand for heterogeneous compute are just a few examples. But today's single-rack scale-up architectures constrain how far domains can grow while meeting the bandwidth and latency requirements of AI workloads.

That's why Eridu helped launch a new workstream within the Open Compute Project focused on disaggregated, multi-rack scale-up architectures.

The initiative is focused on an open, standards-based architecture that extends scale-up across racks while maintaining a single tier of high-performance switching. The goal is to support larger scale-up domains, heterogeneous xPUs, and greater choice in compute and infrastructure components.

We outline the architectural challenges, what is needed to deliver 5000+ GPU scale-up domains, and the principles guiding this initiative at https://t.co/AJ8EPLwbCb.

#AINetworking #AIInfrastructure #ScaleUp

[Delos Data @delosdata]

Decades of experience + the hunger of a first-time founder.

@WenHsiehPhD of @matter_vp on backing @edoe and Dan’s full-stack view of AI data centers.

Watch ⬇️

[Delos Data @delosdata]

GPU utilization in the 30–50% range. That's a lot of idle CapEx.

Christian Gallagher of Socratic on why a better interconnect for inference unlocks so much value at 100K+ GPU scale.

Watch ⬇️

[Delos Data @delosdata]

Missed it at #OCPSummit26?

Watch our CTO Dan Daly and Promode N. of Eridu on Breaking up the Monolith: Multi-rack Scale-up for 5,000+ GPUs.

▶️ https://t.co/AYBE3JMmvZ

[Delos Data @delosdata]

Scale-up is headed past 72 devices to 1,000, then 10,000.

Our CEO @edoe on “In the Arena” from #OCPSummit26: why resiliency matters as much as bandwidth, and why rack scale is a stepping stone to multi-rack AI.

🎧 https://t.co/wPZHyCaVv3

[Prem Jonnalagadda @premosphere]

https://t.co/eSKuw2Ry4G

[Delos Data @delosdata]

Grab a ☕ and learn how to scale up across multiple racks for 5,000+ GPUs.

Dan Daly (Delos CTO) + Promode N. (Eridu CTO) at #OCPSummit26
📍 Today · 8 AM · Room 220B
Can’t make it? https://t.co/8ZEdeE77TK

[Prem Jonnalagadda @premosphere]

@delosdata and @EriduAi on stage at OCP Global Summit. Join us in 220B. #BreakingTheMonolith #OneHop #MultiRack #OCPGlobalSummit @OpenComputePrj

[Delos Data @delosdata]

Sunrise with the Delos team at #OCPSummit26 ☀️

Scaling AI clusters past the rack? That’s the conversation we’re here for.

Request a meeting: https://t.co/8ZEdeE77TK

[Delos Data @delosdata]

NVIDIA’s founders sketched their plans at a Denny’s. Ours started at a dinner in 2024, and yes, we kept the receipt. 🧾

@edoe and Dan Daly on how Delos Data began.

Watch ⬇️

[Pat Gelsinger @PGelsinger]

AI infrastructure is getting bigger and increasingly heterogeneous. GPUs are only part of the story. Networking, memory, CPU, storage, power, cooling and connectivity are all critical to scaling AI.

Some of the biggest breakthroughs may happen far below the model layer.

The infrastructure is the innovation!

https://t.co/2pdmS4nuiZ

[Delos Data @delosdata]

"We invested in Delos because it's the right team focused on the right problem." — @stanreiss , who backed Acacia Communications to a $4.5B exit. 🎥

Thank you for the partnership, Stan.
#fasterinference #strongernetwork #AIinference

[Delos Data @delosdata]

Team. Technology. Market. That's how @PGelsinger decides where to invest, and he explains why Delos Data checks all three.

[Delos Data @delosdata]

We're headed to #OCPSummit26 in San Jose, Oct 12–15! Meet us in our invite-only whisper suite. Send us a note at [email protected]
@OpenComputePrj
#StrongerNetwork #FasterInference

[Eridu @EriduAi]

What would it take to break out of today's limits to the size of scale-up domains, reaching thousands of GPUs in a single-hop domain?

If you're headed to OCP Global Summit in two weeks, join Eridu CTO Promode Nedungadi and @delosdata CTO & Co-Founder Dan Daly as they discuss an open architecture for building multi-rack, single-hop domains of over 5000 GPUs.

They'll present an architecture that shrinks collective-communication time, tightens token-generation tail latency, and lets a frontier model fit entirely inside a single domain.

▶️ "Breaking up the Monolith - An Open Architecture for Multi-rack Scale-up Domain of 5000+ GPUs"
🕗 Wednesday, Oct 14 at 8:00am
📍 Room 220B (SJCC Concourse Level)

#OCPSummit26 #AIInfrastructure

[Sally Ward-Foxton @sallywf]

Thanks Delos Data for the lowdown on your new inference cluster architecture and software which aims to enable GPU (and XPU) scale-up domains up to 1000 and beyond. Buying racks as-is just doesn't offer enough flexibility for some customers:
https://t.co/tSxbmWjZOG

[The Register @TheRegister]

Delos Data offers AI chip startups a fast track to rack scale https://t.co/P93ijSlknM

[Swipeline TR @Swipeline_tr]

Eski Intel çalışanları tarafından kurulan çip girişimi Delos Data, yapay zeka veri merkezlerinde veri aktarımını hızlandıracak çip ve yazılımlar geliştirmek için 100 milyon dolar yatırım aldı. Tura; Matrix Partners, Playground ve Socratic Partners gibi isimler katıldı.

[ImNotHarsh | 📈💸 @imnotharsh]

Delos Data, a startup founded by veterans of Intel Corp $INTC, on Tuesday said it has raised $100 million to develop chips and software that can move data around faster inside AI data centers that are quickly becoming more complex than ever before.

In the early phase of the AI boom, data centers were often made up mostly of Nvidia graphics processing units (GPUs) wired together with proprietary networking technology from Nvidia.

But as the focus of those data centers has shifted from training AI systems to fielding requests from users to have AI "agents" carry out tasks on their behalf, Nvidia rivals such as Cerebras Systems $CBRS and Advanced Micro Devices $AMD have gained traction.

Even Nvidia itself has started offering multiple kinds of computing chips. In response to that new variety of computing chips, Delos Data has designed a complex set of network chips and software with a very simple goal: to help data center owners get all that expensive computing gear communicating as fast as possible, with the flexibility to adapt to whatever mix of computing chips the data center owner might choose, now or in the future.

"We don't know what the next infrastructure architecture is going to be for agentic (AI)," Delos Data Chief Technology Officer and Co-Founder Dan Daly said in an interview. "But we do know that we can provide the quickest, fastest way to move that data around."

Pat Gelsinger, the former CEO of Intel who is now a partner with Playground, a venture capital firm that joined the Delos Data funding round, said computing chips waiting for data remain one of the biggest wastes of money and energy in AI data centers.

"I can have all the computers in the world, and if I can't get these devices communicating effectively with one another, I've got lots of hot hardware, burning watts and sitting here twiddling their thumbs," Gelsinger said.

Other investors in the funding round were Matrix Partners, Playground, Socratic Partners, Capricorn, Matter Venture Partners, IAG and DYNAMIQ.

[Lucas | IA @LucasMestreIA]

A rodada de US$ 100 milhões da Delos Data aponta para um custo que quase nunca aparece na apresentação de um projeto de IA: o tempo em que chips caríssimos ficam esperando dados.

A empresa não está criando outro modelo nem outra GPU. Está trabalhando em chips de rede e software para conectar uma mistura cada vez maior de Nvidia, AMD, Cerebras e outros aceleradores dentro dos data centers.

Isso muda a conta. Comprar mais compute pode piorar o ROI quando a topologia, a movimentação de dados e as filas não acompanham. Você paga energia, refrigeração e depreciação por hardware parado.

Antes de aprovar mais capacidade para agentes de IA, eu pediria três números: utilização real dos aceleradores, tempo perdido em comunicação e custo por tarefa concluída. “Temos mais GPUs” deixou de ser uma resposta financeira.