On August 4, 2026, storage software company Versity made an announcement that, at first glance, sounds like a contradiction in terms: sending data directly from magnetic tape—the oldest digital storage medium, dating back to the 1950s—straight into the high-bandwidth memory of GPUs, today's fastest computing devices. According to CEO Bruce Gilpin, "If you run 50 drives in parallel, you can feed a 20GBps stream into the GPU. And it's remarkably cheap." As the AI industry scrambles to buy up expensive NVMe SSDs, tape—long relegated to warehouse storage—is being pressed back into service.
Why a World That Can't Afford NVMe Is Turning Back to Tape
For a long time, NVMe SSDs and large-capacity DRAM have dominated the storage landscape for AI infrastructure. The prevailing design philosophy held that training data should sit close to the GPU, and that slow storage would only hold back computation. That philosophy was underpinned by abundant capital. As Gilpin himself admits, "Tape hasn't been used in AI because the players involved had no financial constraints."
That premise is now wavering. In his announcement, Gilpin pointed to soaring prices for NVMe and memory, and the now-common reality of multi-month backorders even for 2U servers. The memory supply crunch that has persisted since late 2025 is beginning to delay AI data center construction plans outright. Even with ample budget, the physical components simply aren't available. On top of budget constraints, an "availability" problem is now forcing a rethink of storage design.
Let's review tape's basic specifications here. Under the current LTO-9 standard, a single cartridge offers 18TB of native capacity, and a single drive delivers a native transfer speed of 400MB/s. Compared to the previous generation, LTO-8, capacity increased by 1.5x and speed improved by roughly 11%. A single drive falls far short of an SSD, but a tape library can line up dozens of drives. At 400MB/s per drive, 50 drives running in parallel yields a theoretical 20GBps. The cost per unit of capacity is far below that of HDDs or SSDs, and the media itself can be stored for decades without power.
Control Stays on S3, Data Takes the Back Roads
At the heart of this architecture is a framework called cuObject, which NVIDIA developed as part of its GPUDirect Storage ecosystem. GPUDirect itself is a family of technologies dating back to the early 2010s, providing pathways that move data from NICs or storage into GPU memory without going through the CPU. cuObject applies this same concept to S3-compatible object storage, directly linking GPU memory or system memory with object storage over RDMA transports such as InfiniBand or RoCE.
The trick is splitting S3 requests into two separate paths. Session establishment, authentication, bucket verification, and object metadata all flow through the conventional in-band path via the standard S3 API, just as before. The actual object data, however, travels through an out-of-band path: an RDMA connection running alongside the S3 session. Specifically, a normal S3 GET or PUT request carries RDMA metadata in a custom header called x-amz-rdma-token; the gateway interprets this header and switches only the data transfer over to RDMA. Because the data channel bypasses both the client's CPU and the S3 protocol stack, protocol-processing overhead never touches the transfer itself. The requesting side is managed by cuObjClient, and the gateway side by cuObjServer.
Versity's contribution is connecting its own tape archive management software, ScoutAM (Scale Out Archive Manager), behind this high-speed "front door." The open-source Versity Gateway is stateless by design, translating standard AWS S3 commands into operations against various backends—POSIX file systems, tape via ScoutFS and ScoutAM, or cloud storage. On a PUT, the object arrives at the gateway's memory via RDMA and is then streamed out to the backend. On a GET, the process runs in reverse: data is staged from the backend into the gateway's memory, then sent to the client via RDMA. Any number of GPU-equipped hosts and gateway servers can be deployed, with each pair establishing its own independent RDMA path—allowing the system to scale horizontally.
The tape side has its own optimizations built in. Tape's weakness is sequential access—the time spent rewinding or fast-forwarding to reach the desired data position. ScoutAM reorders requests against large datasets so that offset positions on tape are accessed in an optimal sequence, reducing cartridge mount counts and maximizing streaming performance. Combining optimized read ordering with RDMA transfer, the math suggests that re-reading petabyte-scale datasets can be done within a realistic timeframe.
It's worth noting that the write side of the equation hasn't changed. Ingesting data into the archive remains a steady, conventional ingest process—RDMA's benefits apply mainly to reads. AI training pipelines don't read a dataset just once; they read it repeatedly, every time retraining or reprocessing occurs. That repeated reading is precisely where this architecture shines brightest.
What Sets This Apart from Conventional Setups
| Item | Conventional Tape Archive | Conventional AI-Oriented S3 Storage (NVMe) | Versity GPUDirect Tape Gateway |
|---|---|---|---|
| Data path | Via CPU and S3 protocol stack | Fast, but mostly via CPU | Direct RDMA link between GPU memory and NIC |
| Media cost | Extremely low | High (NVMe, DRAM) | Low (tape-based) |
| Read speed | Slow, retrieval takes time | Fast | Can target 20GBps-class with parallel drives |
| Licensing | Depends on product | Depends on product | Open source, no additional fees |
| Suited for | Cold storage | Hot data during training | Large datasets with repeated reads |
Competitors are moving as well. MinIO already supports S3 over RDMA in AIStor, which it announced in late 2024, claiming zero-copy, kernel-bypass delivery of data directly from AIStor to a GPU's HBM. However, AIStor is a paid, proprietary product, whereas Versity has built RDMA and cuObject support directly into its open-source gateway, with no separate licensing required. Also, in June 2026, ThinkParQ, the company behind BeeGFS, partnered with tape archive vendor GRAU DATA to integrate XtreemStor with BeeGFS. The trend of handling AI and HPC data across both high-performance and long-term storage tiers is becoming an industry-wide direction.
What Remains: Real-World Numbers and Operational Realities
Some caution is warranted here. The 20GBps figure represents a theoretical configuration assuming 50 drives streaming in perfect parallel, and no third-party benchmarks in real-world environments have been published as of this announcement. Tape seek times, cartridge-swap latency, and potential bottlenecks in the GET path—where data is staged once into the gateway's memory—can significantly affect performance depending on how datasets are laid out. It's also unclear who the "large customer currently running trials" that Gilpin references actually is, or what scale of results they've achieved. Additionally, the current implementation of cuObject requires Dynamic Connection transport, meaning compatibility with existing networks must be verified before deployment.
Even so, it's clear that a crack has appeared in the design philosophy underlying AI infrastructure. An industry that has bet everything on computational speed is now running up against component shortages and cost barriers, forcing it to confront an old-yet-new optimization problem: how to read data as fast as possible from cheap, durable media. Whether tape truly settles into its role as the GPU's feeder will be revealed by the real-world performance data yet to come, and by how competing implementations respond.