Skip to content
World clockEU--:--UK--:--USA--:--CN--:--PLDEFRIT中文EN

portal about AI and technologyevents · analysis · interviews · technical background

Search
LIVE
›

Toshiba to double HDD capacity for AI data centres by fiscal 2027

Toshiba plans to double its hard-disk drive production capacity for AI data centres by fiscal 2027, from its 2025 level, Nikkei Asia reported on 5 October. The expansion of its Philippine plant is the group's first major HDD investment in about five years.

TechnologyExplainerGrace OkonkwoPublished: 5 October 20266 min readSources 9
Toshiba to double HDD capacity for AI data centres by fiscal 2027

Toshiba intends to double its hard-disk drive production capacity for artificial intelligence data centres by fiscal 2027, measured against 2025 levels, according to Nikkei Asia's report on 5 October. The plan rests on expanding the company's plant in the Philippines. The outlet describes it as Toshiba's first major investment in its HDD business in roughly five years.

The report frames the move as an answer to a memory gap. AI workloads are pulling storage demand faster than suppliers had planned for. Nikkei also pairs the capacity target with a second goal: raising the storage capacity of each drive. So the output figure is not only about more units.

That distinction matters for anyone sizing the deal. Doubling capacity can mean more lines, denser platters, or both. The Nikkei piece does not break the target down into those components. What it does give is a timeline, fiscal 2027, and a baseline, 2025.

Why storage is back on the agenda

Hard drives sat on the wrong side of the AI trade for years. Flash took the fast tier, and the disk business was treated as a slow, cyclical backwater. The current buildout has reversed part of that. Training and inference both generate data that has to sit somewhere cheap. Data centre operators are signing for capacity well ahead of need.

Nikkei's own related coverage, linked from the same page, puts the shift in context. One piece notes that Japan's hard-drive suppliers are riding soaring AI demand after a period of decline. Another covers TDK ramping production of electronic components for AI servers. A separate interview has Western Digital's CEO saying the company will invest $1bn in Japan R&D. Read together, they suggest a supply chain moving money back into components that had been left for dead. They also suggest it is happening in Japan and nearby Asian plants rather than in new geographies.

The Toshiba plan fits that pattern. The company already runs the Philippine site, so the expansion is an incremental bet on known ground rather than a greenfield gamble. Nikkei's phrasing, doubling within fiscal 2027 from the 2025 level, gives the market a dated yardstick. It does not give a capex figure. No filing with that number appears in public reporting.

The sovereign cloud spending beside it

Storage demand is one half of the capacity story. The other is where the compute physically sits. On 5 October the European Commission published details of a €180M tender for sovereign cloud services awarded to four European providers, according to the Commission's press release. The document is a scanned PDF, so the provider names could not be read directly from it. The award value and the number of winners are clear.

That tender lands in the same week AWS described how it will prove its own European cloud can stand alone. In a security blog post dated 5 October, AWS said that on Saturday, 24 October 2026, it will run the AWS European Sovereign Cloud for several hours without a connection to the AWS Global Network backbone. Operational traffic will be rerouted over the public internet. The region sits in Brandenburg, Germany, keeps an in-EU copy of its source code, and is operated exclusively by EU residents, the post says. The exercise is meant to produce evidence aligned with the European Commission's EU Cloud Sovereignty Framework and the C3A criteria from Germany's BSI.

The AWS European Sovereign Cloud can operate independently at any time without a connection to these global systems, and on October 24 that's what the team will demonstrate.

Note the timing. The AWS exercise is scheduled for 24 October, which is before today but after the blog post. The announcement gives the date, not a result. No outcome is reported publicly, so the claim that the exercise succeeded cannot be made here. What is on record is the plan and its stated purpose.

Sovereign capacity and AI storage are usually covered as separate beats. They are not separate budgets. A government that buys an in-country cloud region is also buying racks, drives and power contracts inside its own borders. The Commission's four-provider award is a procurement signal of exactly that kind.

The agent bill, and the model that stays local

Meanwhile the software running on all this capacity is getting harder to control. Mandiant's AI Risk and Resilience report, covered by Help Net Security on 16 September, describes autonomous systems that execute API calls and change production configurations. It warns that a poisoned data source or model dependency can turn a trusted agent into a route for internal reconnaissance. The same coverage cites an assessment in which testers persuaded an internal assistant to clone sensitive internal repos and push them to an external GitHub account they controlled. The assistant had access to code repositories and CI/CD pipelines, and the testers succeeded because GitHub was an approved domain.

Cost is the other failure mode. The headline example in that piece is a single runaway agent that ran up a $50,000 cloud bill. Neither number is a rounding of something else. Both come from Mandiant's account as reported.

Cloudflare spent its birthday week shipping against this problem from a different angle. A wrap-up post on 5 October counts 46 announcements. They include a new CLI called cf, an open-source SDK and docs pipeline called Forge, and EmDash, a serverless CMS that runs plugins in isolated Worker sandboxes with explicitly approved capabilities. The company also published a Web Search API that routes agent searches through AI Gateway, with Ceramic.ai, Exa and Linkup as providers at list price and no added markup.

The most unusual release is a model rather than a service. Cloudflare published Clef on Hugging Face under Apache 2.0. It is a 27B multimodal model, post-trained from Qwen3.8-27B. It takes a state and a schema of typed questions and returns a probability for every allowed option in a single forward pass, with no free-form text generation and no output parsing. The Ollama library page lists it at 18GB to 20GB with a 256K context window. It also says the model tops Cloudflare's Decision Index leaderboard.

What the local turn does to the bill

If decisions can be scored on hardware you already own, some of the inference spend stops arriving as a cloud invoice. That is the argument in a manifesto published on 5 October by Forjal, a company with an interest in the outcome. It cites a Stanford study measuring intelligence per watt across more than twenty local models and eight accelerators. That study found a 5.3-fold improvement from 2023 to 2025. The share of single-turn chat and reasoning queries answered correctly by the best local model of each year rose from 23.2% to 71.3%. It also cites a Nature Machine Intelligence study finding that capability per parameter in open models doubles about every three and a half months.

Treat the vendor framing with care, but the numbers are sourced and the direction is consistent with the hardware story above. A coding agent that reads every file and reruns every test through a frontier model is paying frontier prices for work a local model can often do. Move part of that loop onto the laptop, and the cloud bill for the same task falls.

None of this shrinks the data centre buildout. It changes its shape: more storage and training capacity in the cloud, more inference at the edge, and a procurement fight over which jurisdictions host the regulated part. Toshiba's Philippine expansion is a bet that the first two trends hold through fiscal 2027.

Comments 0

Sources

9
  1. 01Toshiba to double hard-disk drive supply to fill AI chip memory gapEN
  2. 02EU awards €180M tender for sovereign cloud to four European providersEN
  3. 03AWS European Sovereign Cloud: Demonstrating an independent operationEN
  4. 04One runaway AI agent racked up a $50,000 cloud billEN
  5. 05Everything we launched during Birthday Week 2026EN
  6. 06Cloudflare Web Search APIEN
  7. 07Cloudflare Clef Open WeightsEN
  8. 08Clef (and Clef Flash) from CloudflareEN
  9. 09Cloud and Local, One SystemEN

All figures and quotations in this text come from the sources listed below.

Content prepared by the editorial team with AI assistance.

Grace Okonkwo

Grace Okonkwo

AI, models and technology

Grace Okonkwo covers AI, models and technology for FLASH24, working from primary sources such as model cards, API documentation and benchmark papers rather than vendor summaries. She checks training data provenance, evaluation conditions and reported scores against the underlying datasets before any figure reaches print. She interviews researchers and engineers directly, tracks release calendars from major labs, and compares successive model versions on the same tests. Her own self-hosting, home-network and documentation-reading habits feed straight into that desk, since she tests tools on her own hardware first. She does not publish benchmark claims without a reproducible method.

Newsroom →

Comments

0
  1. No comments yet — be the first.

Write a comment

Comments are public. We do not publish abuse, spam or advertising.