PCIe 6.0 SSDs 2026: Samsung Joins Micron at 28 GB/s

In July 2026, Samsung began mass-producing the PM1763, its first PCIe 6.0 SSD. Five months earlier, in February, Micron did the same thing with the 9650. Two of the three largest NAND makers on the planet are now shipping the fastest storage the specification permits — a shade under 28 gigabytes per second, per drive, sustained. And you, reading a retro-gaming site, cannot buy either one, cannot plug it into anything you own, and will not benefit from it in any measurable way for roughly the rest of the decade.

That is not a complaint. It is the entire story. PCIe 6.0 SSDs are real, they ship in volume, and they were built for a customer that is not you. Here is what actually changed, what the hardware does, and why the gap between the enterprise frontier and the gaming rig under your desk has quietly grown to about five years.

Samsung Joins Micron on the Line

The PM1763 hits mass production

Samsung confirmed in July 2026 that the PM1763 — its first PCIe 6.0 enterprise SSD — had entered mass production. The headline number is 28.4 GB/s sequential read, a touch over double what the company's Gen5-class PM1753 managed. It ships in 4TB, 8TB, and 16TB capacities, uses direct-to-chip liquid cooling, and Samsung's marketing leans on exactly one figure: it can load a 40GB large language model in roughly 1.4 seconds. That is the pitch, in full. Not gaming. Not your Steam library. Cold-starting a neural network fast enough that an idle GPU cluster does not sit there burning money. TechPowerUp logged the production start, and it matters because earlier roadmaps — including the brief this article was commissioned against — had the PM1763 pegged for late-2026 production and 2027 sales. It arrived early.

Micron got there first

Samsung is second. Micron's 9650 series entered mass production in February 2026 and remains, as of this writing, the first PCIe 6.0 SSD anyone shipped in volume — a claim The Register and every outlet that covered it repeated without much dispute. Micron unveiled the drive in July 2025 and was still walking it around Computex 2026 like a trophy. Its numbers: up to 28 GB/s sequential read, 14 GB/s write, and 5.5 million random-read IOPS. Same ballpark as Samsung, arrived first, aimed at the identical customer.

What ‘mass production’ is hiding

Read the phrase mass production and picture a wall of drives at your local Micro Center. Wrong picture. These are E1.S and E3.S rulers — long, thin server modules that slot into hyperscaler chassis, not the M.2 gumsticks in your motherboard. There is no consumer SKU, no MSRP, and no retail channel. The Register put the situation about as plainly as trade press ever does: unless you are building flash storage arrays for AI, you will not have a use for them. That sentence is the thesis of this entire article, and we are going to spend the next 3,000 words explaining why it is correct.

What PCIe 6.0 Actually Changed

64 GT/s and an honest doubling

The PCI-SIG — the industry consortium that owns the standard — ratified PCIe 6.0 in January 2022. The core promise is arithmetic: 64 gigatransfers per second per lane, exactly double PCIe 5.0's 32 GT/s. Scale that to a full x16 slot and you get 256 GB/s of bidirectional bandwidth, a clean 100% increase over Gen5. For a storage drive, which typically receives an x4 allocation, that means a theoretical ceiling around 32 GB/s per direction — which is precisely why the shipping drives all cluster near 28. The specification's history and full generation table are documented in exhausting detail on Wikipedia's PCI Express page, if you want the primary-source cadence.

PAM4, FLIT, and the error-correction tax

Here is the part that matters and that marketing never explains. Gen6 did not double throughput by clocking the same wire twice as fast. It changed how bits ride the wire. Every generation through 5.0 used NRZ signaling — two voltage levels, one bit per symbol. PCIe 6.0 switches to PAM4: four voltage levels, two bits per symbol. You get double the data at the same symbol rate. The catch is that four levels stacked into the same voltage envelope are far easier to misread, so the raw bit-error rate climbs sharply. To compensate, Gen6 makes forward error correction (FEC) mandatory and restructures traffic into fixed-size FLITs — Flow Control Units — instead of variable-length packets. FEC costs latency and a sliver of bandwidth; FLIT mode keeps the correction efficient. This is not a footnote. It is the reason a Gen6 controller is a genuinely hard chip to build, and the reason nobody is dropping one into a $120 gaming SSD.

The x4 math, generation by generation

If you want the whole story in one block, here it is — usable bandwidth for a single x4 NVMe link, the allocation a normal SSD receives:

Usable bandwidth per generation (single x4 NVMe link)
-----------------------------------------------------
Gen     Rate      Signal   Per lane      x4 link
Gen 3    8 GT/s   NRZ      ~0.985 GB/s   ~3.9 GB/s
Gen 4   16 GT/s   NRZ      ~1.969 GB/s   ~7.9 GB/s
Gen 5   32 GT/s   NRZ      ~3.938 GB/s   ~15.8 GB/s
Gen 6   64 GT/s   PAM4     ~7.563 GB/s   ~30.2 GB/s
-----------------------------------------------------
Gen 6 doubles Gen 5 by encoding 2 bits/symbol (PAM4)
plus FLIT + mandatory FEC -- not by clocking higher.
Real drives land ~28 GB/s after overhead.

The takeaway: a Gen6 x4 drive lands near 30 GB/s in theory and about 28 in practice, and it gets there through PAM4 plus FEC overhead, not brute clock speed. That distinction is the entire cost story of this generation.

The Micron 9650, Dissected

The spec sheet, unromanticized

The 9650 comes in two flavors: a read-optimized Pro that scales to 30.72TB and an endurance-tuned Max topping out at 25.6TB. Both hit 28 GB/s sequential read and 14 GB/s write, with 5.5 million random-read IOPS and 900,000 random-write IOPS. Launch reporting framed the gains over Micron's own Gen5 enterprise drives as roughly double the read bandwidth and somewhere between 22% and 67% on the other three headline metrics. Power draw holds at 25 watts — notable not because it is high, but because it is unchanged from a top-end Gen5 enterprise drive while moving twice the data.

G9 NAND and a controller Micron built itself

The flash is Micron's ninth-generation (G9) TLC NAND, running at 3,600 MT/s. The more interesting detail is the controller: Micron designed the ASIC, the DRAM, and the firmware in-house. That is a departure worth flagging, because Micron has historically been content to buy controllers from the merchant market. Doing the whole stack internally is what lets a company chase a first-to-ship title — and it is a quiet signal about how much margin lives in enterprise Gen6, enough to justify taping out your own controller. Hold that thought; the tape-out number is coming, and it is ugly.

Liquid cooling stops being a bragging point

The 9650 is the first Micron SSD where liquid cooling is not a premium option but a near-requirement. Twenty-five watts sounds trivial until you remember the form factor: an E1.S ruler is a sealed sliver of metal with almost no surface area, packed shoulder-to-shoulder with dozens of identical siblings in a chassis with brutal airflow constraints. Alvaro Toledo, Micron's VP and GM for the Core Data Center business, framed the entire product around that pressure: "In an AI driven world where data must move continuously, predictably, and at massive scale, storage performance has become a first order design constraint." Thermals are the tax you pay for that performance — the same reason anyone chasing sustained clocks under load eventually reads up on shaving watts off a component without losing throughput. In the datacenter that means cold plates. On a desktop it means a heatsink and a prayer.

Samsung's PM1763 and the AI Math

28.4 GB/s and the 1.4-second model load

Samsung's spec headline — 28.4 GB/s sequential read — barely edges Micron. The write side lands around 21 GB/s, and the whole drive is rated roughly twice as fast as the prior-generation PM1753. But bandwidth is not the sales hook. The sales hook is that 1.4-second load of a 40GB model. In an inference cluster, the time it takes to swap a model into a GPU's working set is dead time: the accelerator is idle, the power meter is running, the lease is being billed. Halve that and you have a number a CFO understands instantly. Samsung also claims roughly 1.8x the power efficiency of its previous generation, which in a rack where every watt is metered and then cooled is arguably a bigger deal than the raw speed.

The Vera Rubin connection

None of this happens in a vacuum. The PM1763 is explicitly positioned to feed NVIDIA's next-generation Vera Rubin platform — the storage tier that keeps a wall of GPUs from starving on I/O. That is the actual demand signal that pulled the PM1763's production schedule forward from late 2026 into July. When your customer is the AI buildout and your buyer has effectively unlimited capital, timelines compress. The consumer market has never once had that kind of leverage over the storage industry, and it does not have it now.

256TB now, 512TB next

Speed is one axis; density is the other, and Samsung is racing on both. Its published roadmap targets a 256TB PCIe 6.0 SSD in 2026 and a 512TB unit in 2027, both in the EDSFF ruler format, per TechPowerUp's reporting. Micron is on the same treadmill from the other direction with its 6600 ION line — 122TB samples in late 2025, a 245TB monster slated for the first half of 2026. These are capacities that make no sense outside a datacenter and every sense inside one, where a rack's worth of training data needs to sit exactly one PCIe hop from the GPUs.

The Rest of the Field

Silicon Motion's SM8466 ‘MonTitan’

Micron and Samsung ship drives. Silicon Motion sells the controllers other companies build drives around, and its Gen6 part is the SM8466, codenamed MonTitan. Unveiled at the 2025 Future of Memory and Storage summit and fabricated on TSMC's 4nm process, it promises 28 GB/s, up to 7 million IOPS, and support for drives as large as 512TB, with NVMe 2.0-and-up and OCP 2.5 compliance. The important caveat: a controller is not a drive. Announcing silicon in 2025 does not put finished products on shelves, and the first SM8466-based SSDs are not expected to ship until around 2027.

SK Hynix and InnoGrit

SK Hynix, the third member of the NAND oligopoly, confirmed it will join the PCIe 6.0 supply chain before the end of 2026. Its stated targets are aggressive: 25 million IOPS on first samples and a staggering 100 million IOPS on the production version due in 2027 — that latter figure being roughly what NVIDIA has reportedly told the industry it wants from a storage tier. InnoGrit, a smaller controller house, is aiming to launch its own PCIe 6.0 parts within 2026, also targeting up to 25 million IOPS for corporate workloads. The pattern is uniform: every name on the list is chasing the same AI datacenter buyer, and not one of them has mentioned a gaming SKU.

Who ships what, at a glance

The state of play as of mid-2026:

Product	Vendor	Class	Seq Read	Random Read	Max Capacity	Status (mid-2026)
9650 (Pro/Max)	Micron	Drive	28 GB/s	5.5M IOPS	30.72 TB	Mass production, Feb 2026
PM1763	Samsung	Drive	28.4 GB/s	—	16 TB	Mass production, Jul 2026
SM8466 (MonTitan)	Silicon Motion	Controller	28 GB/s	7M IOPS	512 TB	Drives ~2027
Gen6 eSSD	SK Hynix	Drive	—	25M IOPS (sample)	—	Samples by end-2026
Gen6 controller	InnoGrit	Controller	—	25M IOPS	—	Targeting 2026

Why Your Gaming PC Gets Nothing

No CPU, no motherboard, no lanes

Start with the dumbest, most decisive fact: nothing you can buy for a consumer PC in 2026 speaks PCIe 6.0. Not AM5, not LGA1851, not any chipset a normal person can purchase at retail. PCIe 6.0-capable server processors were only expected to arrive around 2026–2027, and consumer platforms trail server platforms by years, every single generation. A drive is useless without a slot that runs it at full rate, and that slot does not exist in your machine. You could not benefit from a Micron 9650 if Micron mailed you one for free — there is nothing in your build to plug it into that would run it above Gen5 speeds anyway.

The $30-40 million question

Even if the platform existed, the economics would not. Silicon Motion CEO Wallace Kou has become the industry's designated realist on this point, and his numbers are brutal: taping out a PCIe 6.0 controller costs somewhere between $30 million and $40 million, against $16–20 million for a Gen5 part. You do not spend $40 million to build a chip for a market that is not asking for it — and per Kou, it is emphatically not asking. "PC OEMs have very little interest in PCIe 6.0 right now," he told the press. "They do not even want to talk about it. AMD and Intel do not want to talk about it." His timeline, delivered at Computex, is the quote this whole beat now orbits: "For consumer? You will not see any PCIe Gen6 [solutions] until 2030." We took that prediction apart in detail when the Micron drive first shipped, in our look at why the 28 GB/s drives ship while gamers wait to 2030, and nothing since has moved the date a single day.

The DirectStorage reality check

Suppose, against all evidence, a Gen6 gaming SSD appeared tomorrow at a sane price. It still would not make your games faster. Real-world game loading is bottlenecked by random 4K reads and CPU-side decompression, not sequential bandwidth, and the industry's answer to that — Microsoft's DirectStorage — remains thinly adopted and barely stresses a Gen4 drive, let alone a Gen5 one. The jump from a Gen4 SSD to a Gen5 SSD is already invisible in load times for the overwhelming majority of titles. Doubling sequential throughput again solves a problem consumers demonstrably do not have. It is the same diminishing-returns trap we keep documenting across PC hardware, from the 28-millisecond gap that makes 240Hz a hard sell over 144Hz to the $1,999 that buys a 30% frame-rate bump nobody asked to pay for. A spec-sheet number is not a benefit until something you actually run can consume it.

How We Got Here: The PCIe Cadence

From 2.5 GT/s to 64 in two decades

PCI Express is old. The 1.0 specification landed in 2003 at 2.5 GT/s per lane, and every version since has roughly doubled the per-lane rate: 5.0 GT/s (2.0, in 2007), 8 GT/s (3.0, 2010), 16 GT/s (4.0, 2017), 32 GT/s (5.0, 2019), and 64 GT/s (6.0, 2022). Doubling per-lane bandwidth every few years is a remarkable engineering cadence — and it is also the source of the problem, because the specification consistently outruns the silicon and the platforms that have to implement it.

The consumer lag is structural, not accidental

Look at when each generation actually reached a consumer SSD you could buy, and a pattern emerges:

Generation	Spec Ratified	Per-Lane Rate	x16 Bidirectional	First Consumer SSD
PCIe 3.0	2010	8 GT/s	32 GB/s	~2015
PCIe 4.0	2017	16 GT/s	64 GB/s	2019
PCIe 5.0	2019	32 GT/s	128 GB/s	2023
PCIe 6.0	Jan 2022	64 GT/s	256 GB/s	~2030 (projected)
PCIe 7.0	Jun 2025	128 GT/s	512 GB/s	Mid-2030s+

The lag between a ratified spec and a shippable consumer drive has run anywhere from three to eight years, and for Gen6 it is tracking toward the high end of that range. This is not manufacturers dragging their feet — it is the honest cost of PAM4 signaling, mandatory FEC, controller complexity, and the plain fact that consumer platforms wait for the enterprise market to amortize the R&D first. Enterprise pays the tape-out bill; consumers inherit the technology once the volume math finally works.

PCIe 7.0 already exists on paper

If you want proof that the spec outruns reality, consider that the PCI-SIG finalized PCIe 7.0 in June 2025 — 128 GT/s per lane, 512 GB/s bidirectional across x16 — while Gen6 SSDs were still months away from mass production. The Register caught PCI-SIG president Al Yanes hedging on what comes after: "We are hoping to double again, but I do not want to make any definitive claims at the moment." Industry consensus already holds that consumer PCIe 7.0 is more than a decade out. The paper standard is two full generations ahead of the fastest thing you can buy for a desktop. That gap is the entire shape of this market.

What the Builders Actually Say

Kou draws the line at 2030

No single quote defines this beat like Wallace Kou's. Speaking as the CEO of Silicon Motion — the largest merchant SSD-controller vendor on Earth, and therefore a company with every incentive to hype a new interface — he instead told PCGamesN the opposite: "For consumer? You will not see any PCIe Gen6 [solutions] until 2030." He was equally candid about his own book, noting that Silicon Motion dominates the current generation and is content to keep milking it: "We dominate PCIe 5.0, both 8-channel and 4-channel controllers. For the next four years, we will be in a comfortable position to continue growing in the client market." When the controller vendor says the party is not starting until 2030, the party is not starting until 2030. Tom's Hardware logged the same remarks alongside those tape-out cost figures.

Micron sells the AI necessity

From the vendor actually shipping, Alvaro Toledo's framing is worth repeating because it is the honest sales pitch stripped of spin: "In an AI driven world where data must move continuously, predictably, and at massive scale, storage performance has become a first order design constraint." Note the words that are absent — no "gamer," no "creator," no "prosumer." The value proposition is entirely about feeding models, and Micron is not pretending otherwise.

The blunt version

And then there is the trade press cutting through the marketing entirely. The Register's verdict on the Micron 9650, published the week it hit mass production, is the sentence to tape above your desk: unless you are building flash storage arrays for AI, you will not have a use for them. Two shipping drives, a queue of vendors behind them, one customer — and the customer is a datacenter.

The Next 6 to 12 Months

What will happen

Nothing here requires a crystal ball — just extrapolating stated roadmaps through mid-2027:

1. SK Hynix ships Gen6 samples before New Year. The company has committed to entering the supply chain by the end of 2026 with 25-million-IOPS samples, and a 100-million-IOPS production part targeted for 2027. Expect the sample announcement inside this window.
2. Samsung ramps density, not accessibility. The 256TB Gen6 drive lands in 2026 as promised, and the 512TB unit gets pulled into 2027 conversation. None of it approaches a consumer channel. The actual news will be more Vera Rubin design wins.
3. Silicon Motion's SM8466 drives finally appear — for servers. The controller was 2025's announcement; the first real drives built on it surface around early 2027, still enterprise-only. Its client-oriented Gen6 controller (25+ GB/s, 3.5M IOPS) stays penciled for 2028–2029.
4. Zero consumer motherboards gain PCIe 6.0. No AM5 refresh, no Intel platform, no chipset will expose Gen6 to a desktop in this window. The consumer story does not move; Kou's 2030 date holds.
5. PCIe 5.0 stays the consumer ceiling — and prices stay ugly for other reasons. Gen5 SSDs remain the fastest thing you can actually buy, and the ongoing NAND and DRAM supply crunch keeps upward pressure on all SSD pricing regardless of interface. If anything on your shopping list gets more expensive, it will be Gen5, and it will have nothing to do with Gen6.

What almost certainly will not

No consumer Gen6 SSD. No Gen6 gaming motherboard. No pricing announcement for either the 9650 or the PM1763, because enterprise drives sell on negotiated contracts, not MSRPs. And no benchmark showing a Gen6 drive loading a game meaningfully faster than a Gen5 one — because that benchmark cannot exist on hardware nobody can assemble.

The Verdict

For the datacenter: genuinely a big deal

If you operate an AI training or inference cluster, the arrival of two shipping PCIe 6.0 drives from Micron and Samsung, with SK Hynix and Silicon Motion right behind them, is a real and useful step. Doubling storage bandwidth to 28 GB/s, cutting model-load times to seconds, and improving power efficiency by roughly 1.8x are meaningful gains when your bottleneck is feeding GPUs and your hard constraint is the electricity bill. This technology solves a problem that a specific, extremely wealthy customer actually has.

For the gamer: a spectator sport

For everyone reading this on a machine they built themselves, PCIe 6.0 in 2026 is a spectator sport. You cannot buy the drives, you have no platform to run them, and even if both facts flipped tomorrow, your games would not load faster. It is the same lesson the last few years of PC hardware keep teaching: the number on the box is not the experience at your desk. The $300 monitor-module tax that finally collapsed died for exactly this reason — buyers eventually stop paying premiums for differences they cannot perceive. Gen6 will reach you around 2030, at a sane price, once it is thoroughly boring. That is not a tragedy. That is how every PCIe generation has always worked.

The bottom line

Two PCIe 6.0 SSDs now ship in volume. Both are enterprise rulers built for AI datacenters, both need liquid cooling, neither has a consumer price, and the man who makes most of the world's SSD controllers says your turn comes in 2030. The frontier is real. It is just very far from your desk — and it was never walking in your direction.