ibl.ai Agentic AI Blog

Insights on building and deploying agentic AI systems. Our blog covers AI agent architectures, LLM infrastructure, MCP servers, enterprise deployment strategies, and real-world implementation guides. Whether you are a developer building AI agents, a CTO evaluating agentic platforms, or a technical leader driving AI adoption, you will find practical guidance here.

Topics We Cover

Featured Research and Reports

We analyze key research from leading institutions and labs including Google DeepMind, Anthropic, OpenAI, Meta AI, McKinsey, and the World Economic Forum. Our content includes detailed analysis of reports on AI agents, foundation models, and enterprise AI strategy.

For Technical Leaders

CTOs, engineering leads, and AI architects turn to our blog for guidance on agent orchestration, model evaluation, infrastructure planning, and building production-ready AI systems. We provide frameworks for responsible AI deployment that balance capability with safety and reliability.

Back to Blog

Phison's AI SSD: What an NPU Inside the Drive Changes

Miguel AmigotNovember 22, 2026
Premium

At SC26 on November 17, 2026 Phison previewed a computational-storage SSD with an integrated NPU and 64 GB of 130 GB/s memory. It is still in development, with no product name or sample date, and it moves filtering to the data rather than replacing GPUs.

The Short Answer

Phison did not ship an AI SSD. At SC26 on November 17, 2026 it previewed an unnamed computational-storage design: an NPU and 64 GB of 130 GB/s memory inside the drive, running PyTorch, vLLM, OpenVINO and llama.cpp, with no sample date. It is a data-movement play, not a GPU replacement. On ibl.ai you own all the code and the data, wherever inference runs.

The viral version says "no GPU cluster, no data movement." Phison's own release claims fewer data transfers for suitable workloads, which is a smaller and more defensible statement.

What exactly did Phison announce at SC26?

Phison announced three storage technologies in a release dated Chicago, November 17, 2026, timed to Supercomputing 2026. Only one of them is the "AI inside the SSD" story.

The computational-storage design pairs an integrated NPU, rated by Phison at 36,800 DMIPS, with 64 GB of memory at 130 GB/s.

Phison says it supports PyTorch, vLLM, OpenVINO and llama.cpp, and claims a potential 10–100x reduction in data transfers for applicable workloads, with lower host CPU use.

The other two announcements were a PCIe Gen6 enterprise controller, the X3 (PS5303), targeting up to 28 GB/s and 6.8 million IOPS, and a "Mixed Media" drive that puts pSLC and QLC flash behind one 16-channel controller.

What the viral post says What Phison's release says
Phison "just put" an AI processor in the SSD An architecture previewed at SC26; "many of the products and features mentioned are still in development"
An AI processor + 64GB memory Accurate: an integrated NPU and 64 GB of memory at 130 GB/s
No data movement A potential 10–100x reduction in data transfers, for applicable workloads
No GPU cluster Not claimed. The release describes fewer data transfers and lower host CPU use

Two of the four claims hold up. The two that do not are the ones a hardware budget would be built on.

Is Phison's AI SSD a shipping product, and is it the same thing as aiDAPTIV+?

No on both counts. Tom's Hardware's November 20, 2026 write-up notes the design has no product name, no named controller and no sample date.

The same report lists what Phison left undisclosed: the memory type, form factor, power draw, a TOPS rating, how a host reaches the drive's compute, and which model sizes it supports.

It also flags that 36,800 DMIPS is a decades-old integer CPU benchmark, not an AI throughput figure, and may describe the companion CPU cores rather than the NPU. Treat it as saying little about AI performance.

It is also distinct from Phison's aiDAPTIV+, the product many readers will remember.

aiDAPTIV+, now sold as Pascari aiDAPTIV, uses flash and system memory to extend the memory a GPU can use, so larger models run than GPU memory alone would hold. The SC26 design does the opposite: it runs compute on the drive itself.

Computational storage is not new either. Tom's Hardware cites Samsung's SmartSSD and ScaleFlux as earlier designs in the same category.

Can a 64 GB NPU inside an SSD run a large language model?

It can hold a small or quantized one, but memory bandwidth caps how fast it generates. For a dense model, producing each token means reading every weight once, so tokens per second cannot exceed bandwidth divided by the model's size in bytes.

That ceiling is simple arithmetic, and it is ours, not Phison's. Phison published no inference throughput. Real throughput lands below the ceiling once compute, the KV cache and overhead are counted.

Dense model, 4-bit weights Weights Ceiling at 130 GB/s (Phison drive) Ceiling at 3.35 TB/s (H100 SXM)
8B parameters ~4 GB ~32 tokens/s ~840 tokens/s
70B parameters ~35 GB ~3.7 tokens/s ~96 tokens/s

The H100 figure is NVIDIA's published 3.35 TB/s for the H100 SXM. The drive's 130 GB/s is about 4% of that.

So a 70B model fits in 64 GB at 4-bit, but a ceiling of under four tokens a second is not an interactive chat service. "No GPU cluster" holds only for work that was never going to need one.

What workloads does near-data inference actually suit?

Work where the expensive part is moving data, not generating text. Phison's own framing is a reduction in data transferred to the host, which points at filtering, classification, embedding and pre-processing over large stored datasets.

The pattern is a scan that would otherwise ship terabytes across the bus so a host can discard most of it. If a small model on the drive can tag, filter or embed records in place, only the useful slice travels, and that is where a 10–100x transfer reduction is plausible.

A retrieval pipeline is the natural fit: embedding documents where they sit, then sending vectors rather than raw files to the serving tier.

The generating model still runs where bandwidth is high. For that tier, the self-hosted LLM serving stack β€” vLLM or TGI on dedicated GPUs β€” stays the answer.

This is the same honest boundary as NVIDIA's PAIR, which routes requests rather than pooling GPU memory: local and near-data inference change where work runs, not how large a model one device can serve quickly.

Does moving AI compute to the data change enterprise AI cost?

It changes the hardware line, and it makes a pricing argument plain. Compute inside a drive you already own has no seat count in it. Its cost is the hardware and the electricity, and it does not move when you hire.

Per-seat AI licensing bills headcount regardless of use. Microsoft 365 Copilot is listed at $30 per user per month on an annual commitment, which is $720,000 a year at 2,000 employees whether those seats send one request a day or none.

Per-seat is structurally the wrong shape for AI at scale, not one pricing option among several. The more inference migrates into hardware the enterprise owns, from GPUs to NPUs to storage, the less a headcount-multiplied bill describes what is actually being consumed.

Does in-storage AI make enterprise AI more sovereign?

Only if the drive sits on infrastructure you control. Near-data processing shortens the path data travels inside one server; it does not decide whose server that is.

A drive with an NPU in a hyperscaler's rack is still the hyperscaler's hardware. A drive in your own data centre was already inside your perimeter before it gained an NPU.

What on-drive compute adds is one more place where software runs against your data, and that software has to be governed like any other. Phison has not said how a host reaches the drive's compute, which is exactly the detail a security review will ask for.

Sovereignty is decided by who owns the stack and where it runs. Storage-level compute is a performance feature that a sovereign deployment can use, not a substitute for one.

How does ibl.ai fit near-data and on-device inference?

On ibl.ai you own all the code and the data. The platform runs on your infrastructure with full source code, is model-agnostic across any LLM, has no per-seat pricing, and can deploy anywhere: your own cloud, on-premise, GovCloud, or a fully air-gapped network.

Inference placement is already an operator decision on the platform. The September 4, 2026 platform update shipped an on-device LLM provider picker, so an operator chooses where a given agent's inference runs, alongside hosted providers.

When computational storage reaches products with published interfaces, it becomes one more placement option under the same rule: the data stays where you put it, on hardware you control. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY.

Related reading: NVIDIA's PAIR is a router, not an inference cluster and self-hosted LLM providers: Ollama vs vLLM vs TGI vs LocalAI.

Sources: the NPU, 64 GB / 130 GB/s memory, framework list, 10–100x claim, X3 and Mixed Media details and the in-development disclaimer from Phison's SC26 release; the missing product name and sample date, the undisclosed details, the DMIPS caveat and the SmartSSD and ScaleFlux precedents from Tom's Hardware; aiDAPTIV+ from Phison; H100 bandwidth from NVIDIA; Copilot pricing from Microsoft. Token-rate ceilings are our arithmetic, not vendor figures.

Why does owning the AI stack matter?

ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing β€” so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.

  • You own all the code and the data

    Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform β€” the stack itself is yours.

  • Model-agnostic

    Run any LLM β€” Claude, GPT, Gemini, Llama, Command, or your own fine-tune β€” and switch providers without rewriting the platform.

  • No per-seat pricing

    Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ.

  • Deploy anywhere

    Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity.

1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University.

ibl.ai is family-owned and operated from New York, NY β€” a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.

Related Articles

Self-Hosted AI for Universities: FERPA-Safe by Design

Self-hosted AI for universities means the runtime executes inside infrastructure the campus controls β€” FERPA-protected student records never leave the institution boundary. The deployment options, the workloads, the cost math, and why this becomes the default endpoint for any serious campus AI program.

Mikel AmigotJune 8, 2026

Self-Hosted AI Agent Platform You Own: All the Code, All the Data

A self-hosted AI agent platform you own = the source code, the runtime, the model, and the data inside your infrastructure. ibl.ai is the platform: open-source runtime, perpetual license, any LLM, deploy anywhere, no per-seat pricing.

Blanca AmigotJune 1, 2026

Self-Hosted AI for Hospitals and Health Systems: The Deployment That Survives Audit

Self-hosted AI for hospitals and health systems means the runtime executes inside your existing HIPAA-covered environment β€” PHI never traverses a third-party cloud. The deployment options, the workloads, the cost math, and why this becomes the default endpoint for any serious clinical AI program.

Mikel AmigotJune 1, 2026

ibl.ai on Microsoft Surface Copilot+ PCs: Local AI Tutoring Powered by the NPU

ibl.ai runs directly on Microsoft Surface Copilot+ PCs, using the built-in Neural Processing Unit (NPU) to deliver real-time AI tutoring and content tools without requiring a cloud connection. Students get instant, on-device mentoring; faculty get powerful authoring tools; and institutions keep every byte of data local.

Miguel AmigotFebruary 13, 2026

See the ibl.ai AI Operating System in Action

Discover how leading universities and organizations are transforming education with the ibl.ai AI Operating System. Explore real-world implementations from Harvard, MIT, Stanford, and users from 400+ institutions worldwide.

View Case Studies
Work with our team

Pilots, deployment, and full ownership

Every engagement is scoped to the work: you integrate ibl.ai with your own data, deploy it on your own infrastructure, and the engineering hours scale with the scope β€” so the price tracks the scope, not your headcount. Most customers keep us on afterwards for platform updates, new agents and integrations, and ongoing services.

Start here

Pilot

from $15K

fixed scope Β· fixed timeline

A time-boxed proof of value on your real data β€” not a slide deck.

Best for: Teams that want to see ibl.ai working before committing.

  • Deployed on your infrastructure or our cloud
  • 1–2 production agents wired to a slice of your data
  • One integration (LMS / SIS / SSO / data source)
  • Weekly working sessions with our engineers
  • Pilot fee credits toward a full engagement
Scope a pilot
Most common

Integration & Deployment

$25K – $80K

one-time fee Β· scoped to the work

Full deployment integrated with your data and systems. Engineering hours scale with scope.

Best for: Organizations rolling ibl.ai out across a department, campus, or business unit.

  • Platform deployed in your VPC, on-prem, or air-gapped
  • Integrated with your data + identity (SSO / SAML)
  • Multiple custom agents built to your workflows
  • Engineering hours proportional to scope
  • You own the data Β· run any LLM you choose
Plan a deployment
Full ownership

Codebase Transfer + Custom AI Engineering

Custom quote

perpetual license Β· you own the stack

We transfer the full source code. You own and self-host the entire platform β€” outright.

Best for: Organizations and enterprises that benefit from perpetual ownership and sovereignty.

  • Complete source-code transfer + perpetual license
  • Dedicated AI engineering team on your roadmap
  • Custom agents, models, and integrations to spec
  • Air-gapped capable Β· zero vendor lock-in
  • Family-owned, New York–based long-term partner
Talk about ownership
You own the code and data Run any LLM β€” Claude, GPT, Gemini, Llama Family-owned & operated from New York, NY