Four years after launch, DirectStorage is still waiting for the one game that proves it matters


Every few years, the PC gaming world gets handed a piece of tech that’s supposed to change everything, and then it just kind of … sits there. Ray tracing took the better part of a decade. VR is still figuring itself out. And then there’s DirectStorage, which Microsoft announced back in 2020 but has barely been used since.

And it’s a shame, because the idea behind it is genuinely great: your SSD is already fast enough to make loading screens disappear, but Windows is standing in the way.

Your SSD was never the bottleneck, Windows was

The way games load files was designed for spinning platters

NVMe SSDs are stupidly fast; they can move multiple gigabytes per second. But the way Windows traditionally hands files to a game was built back when storage meant a mechanical drive with a physical arm swinging around inside it. Microsoft has openly said the standard file I/O APIs are more than 30 years old and basically unchanged.

Modern games don’t ask for one giant file anymore. They constantly stream thousands of tiny assets as you move through a world, which generates an absolutely enormous pile of small requests. Old-school storage APIs choke on that pattern, and all that overhead lands on your CPU. If you want to go deeper on why the number on the SSD box lies to you, sequential read speed isn’t the spec that actually matters. Windows itself has historically added latency by treating NVMe drives like ancient SCSI devices.

On top of the request overhead, there’s decompression. Game assets ship compressed to keep install sizes manageable, so something has to unpack them. Traditionally, that’s your CPU, chewing through gigabytes while the GPU taps its foot, waiting.

What DirectStorage actually does

Fewer detours, and the GPU does the unpacking

Closeup of graphics card in ITX case Credit: Ismar Hrnjicevic / How-To Geek

DirectStorage attacks both of these problems. First, it lets games talk to a fast NVMe drive with way less operating system overhead, batching requests instead of politely queuing them one at a time. Second, and this is the flashy bit, it can hand the decompression job to your GPU instead of your CPU.

That sounds strange until you remember what a GPU is: thousands of small cores that are phenomenal at doing the same repetitive thing in parallel. Unpacking compressed data is exactly that kind of work. DirectStorage 1.1 introduced a compression format called GDeflate, developed by Nvidia and contributed as an open standard, built specifically to be chewed up in parallel, and the data stays compressed all the way until it hits VRAM. Your CPU still handles the I/O requests, but it no longer does all the heavy lifting.

There are plenty of requirements to meet (you’ll need to be on Windows 10 or 11, using an NVMe drive and a DirectX 12 GPU, and playing a supported game), but if all that lines up, the tech will switch on automatically and do its thing.

The Xbox got the better version of this

Consoles have a dedicated chip for it, and your PC doesn’t

An Xbox controller in front of Cyberpunk 2077. Credit: Nick Lewis/How-To Geek

DirectStorage started life on the Xbox Series X and S, where it’s just one piece of something Microsoft calls the Velocity Architecture. The other pieces are a custom 1TB NVMe SSD, Sampler Feedback Streaming (which loads only the slivers of a texture the GPU can actually see, for roughly 2.5x the effective memory and I/O), and, crucially, a dedicated hardware decompression block.

On Xbox, unpacking assets happens on purpose-built silicon that does nothing else. But on your PC, there is no such chip. If the GPU is decompressing, it’s spending compute and power budget it could have spent on frames. Every second of loading you save is potentially borrowed from your frame rate, and that tradeoff is basically the entire story of DirectStorage on PC.

The games that use it, and how that went

A very short list, and a couple of famous faceplants

Aloy being approached by robot dinosaurs in Horizon Forbidden West. Credit: Sony

Four years after launch, the roll call is thin: Forspoken got there first in January 2023, with Ratchet & Clank: Rift Apart and Forza Motorsport following later that year, then Horizon Forbidden West in 2024. Marvel’s Spider-Man 2 joined in early 2025. Plenty of other games (Diablo IV, New World, EA Sports FC 24) ship the DirectStorage files without necessarily using the tech in any meaningful way, which tells you how much of this is theater.

For the games that do or did use DirectStorage, not everything went smoothly. Forspoken loaded in one or two seconds, which is great, except testing showed a screaming Gen5 SSD barely beat a sad SATA drive, so the tech wasn’t really being pushed. Ratchet & Clank was worse. ComputerBase found that deleting two DirectStorage DLLs from the game folder gave an RTX 4080 about 10% more average FPS at 4K and a wild 26% improvement in 1% lows. AMD cards were completely unaffected. Deleting the files didn’t even hurt load times.

Arguably the most damning example, though, was Horizon Forbidden West, which shipped with DirectStorage but deliberately did not use GPU decompression, saying the API was too restrictive and that GPU decompression caused driver scheduling issues they couldn’t work around. When the porting studio that knows this tech best opts out of the headline feature, that’s a signal.

So why has nobody adopted it?

Because it’s a rebuild, not a checkbox

Four different NVMe SSDs installed in the TerraMaster F4 SSD NAS. Credit: Patrick Campanale / How-To Geek

Turning on DirectStorage properly means rethinking how your game packages, compresses, and streams every asset it owns. That’s a serious engineering investment in something largely unproven, and studios have limited budgets.

Then there’s the PC problem. Xbox devs know exactly what drive is in the box. But with PCs, users might be running a Gen5 monster, a tired SATA SSD, or even a hard drive. You can’t design a game with zero loading screens while still supporting all of that, and locking people out costs sales. Add the fact that most games are GPU-bound anyway, and the pitch becomes “give up frames to save seconds,” which isn’t very convincing.


Should you care about DirectStorage?

DirectStorage really has nothing to do with us as consumers; there’s no product to buy into. Any PC with an NVMe SSD will be able to use it, if game developers ever decide it’s worth designing for. With the release of DirectStorage 1.4 in 2026, perhaps the narrative will begin to shift, but all we can do is wait and see.

If you’re running old tech at the moment, please do not buy an expensive Gen5 SSD because of DirectStorage, since a faster drive barely moves the needle in games anyway, and it is not worth it at all with SSD prices being what they are right now.

The Samsung 9100 PRO NVMe SSD.

7/10

Storage capacity

1TB, 2TB, 4TB, 8TB

You don’t need the fastest SSD in the world for gaming, but if you feel like a splurge, the Samsung 9100 Pro is super fast and premium-quality.




Source link

Leave a Reply

Subscribe to Our Newsletter

Get our latest articles delivered straight to your inbox. No spam, we promise.

Recent Reviews


TL;DR

India debates sovereign AI after the US forced Anthropic to kill Fable 5, with proposals for a $5B fund and calls to embrace open-source models.

When the US government ordered Anthropic to shut down Fable 5 and Mythos 5 on 12 June, the export control directive was aimed at restricting foreign nationals from accessing America’s most capable AI. In India, Anthropic’s second-largest market, it landed as a warning shot about what happens when your AI infrastructure runs on someone else’s politics.

The suspension cut off Indian developers and enterprises from Claude’s most advanced models overnight. India’s Claude run-rate revenue had doubled since October 2025, and Tata Consultancy Services had announced a partnership just one day earlier, on 11 June, to train 50,000 employees on Claude and build a dedicated Anthropic business unit. That deal is now in limbo.

The timing has turned what was already a simmering debate about AI sovereignty into a full strategic reckoning. Proposals that sounded ambitious a week ago now sound urgent.

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

Mohandas Pai, former Infosys CFO and one of India’s most prominent tech investors, has called for a ₹50,000 crore (roughly $5 billion) annual sovereign AI fund. He has also proposed a ₹2 lakh crore (approximately $21 billion) credit guarantee to finance cloud infrastructure, hardware procurement, and semiconductor development. The figures dwarf the government’s existing commitment.

India approved its IndiaAI Mission in March 2024 with a budget of ₹10,372 crore, approximately $1.25 billion. The programme has deployed around 38,000 GPUs so far. Pai’s proposal would quadruple annual spending and add a credit backstop an order of magnitude larger.

Sridhar Vembu, the founder of Zoho, has gone further. He argued that India should embrace smaller and open-source models, including Chinese ones, rather than depend on American frontier systems that can be switched off by executive order. “Technology is the ultimate weapon,” Vembu said. “Globalization is dead and Bharat must find her own way ahead.

The argument has teeth because the suspension demonstrated exactly the vulnerability Vembu is describing. Amazon’s CEO reportedly triggered the government crackdown by telling Treasury Secretary Scott Bessent that researchers had used Fable 5 to obtain information that could be used in cyberattacks. Anthropic called the action disproportionate, but compliance was immediate and global.

Policy expert Prasanto Roy put it bluntly: “American AI models are bound to American geopolitics.” For Indian enterprises that had built workflows around Claude, the lesson was that access to frontier AI is a privilege that can be revoked without notice, without consultation, and without regard for the commercial relationships it disrupts.

The Indian startup ecosystem is already adapting. Sarvam, a Bengaluru-based AI company, released 30-billion and 105-billion parameter open-source models at the India AI Impact Summit in 2026. Krutrim, founded by Ola’s Bhavish Aggarwal, has pivoted from building foundational models to providing cloud and AI infrastructure services, reporting ₹3 billion in revenue for fiscal year 2026.

Neither company is close to matching the capabilities of Fable 5 or Mythos 5. But the argument for sovereign AI was never about matching frontier performance immediately. It is about ensuring that the floor does not fall out when Washington makes a unilateral decision about who gets to use which models.

Aakrit Vaish, founder of the AI startup Activate, said the suspension “completely changes things” for the sovereign AI debate. Vijay Rayapati, CEO of Atomicwork, raised concerns about what the precedent means for Indian companies with multi-country teams that depend on American AI providers. If the US can shut off model access to enforce export controls, any country that relies on American AI is one policy decision away from disruption.

Not everyone agrees that India needs to build its own frontier models. Hemant Mohapatra, a partner at Lightspeed Venture Partners, argued that talent and compute access matter more than capital for building competitive AI. India has the engineering workforce, but the compute gap is significant, and closing it requires either massive domestic investment or continued access to foreign cloud infrastructure.

Anthropic opened a Bengaluru office as part of its India expansion, and the TCS partnership was designed to be a cornerstone of its enterprise strategy in the country. Whether those plans survive the suspension intact depends on how quickly Anthropic can restore access and whether Indian enterprises still trust a provider whose most capable models can vanish overnight.

The broader pattern is unmistakable. The US has spent four years tightening controls on AI technology, from chip export restrictions to model-level interventions. Each escalation pushes more countries toward the conclusion that dependence on American AI infrastructure carries political risk. India, with its 1.4 billion people and rapidly growing technology sector, is now asking whether it can afford that risk, and what it would cost to eliminate it.

The Opendoor layoffs in June 2026, which shut the company’s India office and affected roughly 250 employees, added another dimension. CEO Kaz Nejatian cited AI-native teams as the reason, suggesting that some US companies are using AI to reduce their reliance on Indian engineering talent at the same time that India is debating its reliance on American AI. The relationship is becoming less complementary and more competitive.

For now, the sovereign AI proposals remain proposals. Pai’s fund has no legislative vehicle, Vembu’s call for open-source adoption has no coordinated policy framework, and the IndiaAI Mission’s GPU deployment is still in early stages.

But the Anthropic suspension has done something that years of policy papers and conference speeches could not: it has given the sovereign AI movement a concrete, recent, and viscerally felt example of why dependence on foreign AI is a strategic liability. The debate is no longer theoretical.



Source link