Farewell to DDR5 Monopoly? What Optical Fiber Memory Actually Means

When legendary programmer John Carmack — the father of the game engine, FPS pioneer, and serial boundary-crosser from gaming to rockets to VR to AI — posts a bold idea on X, the internet listens. His recent suggestion that optical fiber could serve as a radical new form of computer memory has triggered a wave of breathless coverage declaring the death of DDR5 and the dawn of photonic storage. The reality is far more nuanced, and far more interesting.

The core concept is real. The mathematics check out. Top figures in tech, including Elon Musk, have signalled approval. But the popular-science circuit has been selectively quoting the advantages while quietly omitting the structural limitations — creating a wildly distorted picture of what this technology can actually do.

Not a New Memory — An Ancient Idea Reborn in Photons

The official name for what Carmack describes is Fiber Delay Line Storage (FDL). Despite the futuristic framing, delay-line storage is a concept older than the transistor. Early vacuum-tube computers used acoustic delay lines — bouncing sound waves through mercury columns to hold data temporarily. Today’s proposal simply replaces mercury with single-mode optical fibre, and sound with light.

The operating principle is fundamentally different from the RAM you know. There are no storage cells. There is no static data retention. Information exists only because optical signals are continuously looping through a fibre ring, their presence constituting a transient “hold.”

That distinction — cabinet versus conveyor belt — is not an implementation detail. It is the technology’s defining characteristic, and the root of both its extraordinary strengths and its equally extraordinary limitations.

Where the Hype Is Justified: The Numbers Are Genuinely Extraordinary

Carmack’s back-of-envelope calculation is, by all technical assessments, correct. A single-mode fibre with 256 Tb/s of bandwidth running a 200 km loop produces exactly 1 millisecond of delay. The resulting instantaneous data capacity is 32 GB, with an effective bandwidth of 32 TB/s.

The bandwidth advantage is real — four times what the best DDR5 can offer. So too are two secondary benefits: because optical signals are passive transmissions that require no refresh cycles, static power consumption approaches zero in theory (DRAM clusters lose over 30% of their power budget to constant refresh operations alone). And because FDL relies on fibre physics rather than lithography, it entirely bypasses the semiconductor process-node arms race, making the “leap-frog” narrative at least logically coherent.

Five Structural Flaws That Cannot Be Engineered Away

Every one of the following limitations is not a temporary engineering challenge awaiting a clever fix. They are consequences of the underlying physics, baked in by the same principles that produce the impressive bandwidth numbers.

• 1
  Latency: a 10,000× chasm A full loop through the fibre takes 1 millisecond. A CPU’s L2 cache operates at roughly 100 nanoseconds. That is a ten-thousand-fold difference. For general computing, the CPU would sit idle for an eternity of processor time waiting for data to come around the loop.
• 2
  Sequential access only — no random addressing Modern operating systems, applications, and games depend entirely on the ability to fetch data from arbitrary memory addresses instantly. FDL is first-in, first-out. You cannot skip the queue. Think cassette tape, not SSD.
• 3
  Signals decay — active hardware required throughout Optical signals attenuate over 200 km. Maintaining signal integrity demands optical amplifiers and DSP chips distributed across the loop, eroding the power-saving advantage and adding complexity and failure points.
• 4
  Capacity and latency are inseparably coupled More storage means longer fibre, which means higher latency. Scaling to terabyte-class AI models would require thousands of kilometres of fibre, at which point the latency becomes commercially absurd.
• 5
  Physical scale and engineering complexity 200 km of fibre, however tightly wound, is not going into a DIMM slot, a smartphone, or a rack unit. Temperature stability, interference shielding, and fault tolerance at fibre-loop scale present serious industrial engineering challenges.

Not a Replacement — a Specialised Complement

None of this makes FDL pseudoscience or a failed idea. It makes it a highly specific tool being misidentified as a universal one. The technology’s actual destiny lies in the handful of computing contexts where its weaknesses are irrelevant and its strengths are transformative.

AI workloads are, by nature, sequential and streaming — vast rivers of matrix data flowing in predictable order through a compute pipeline. That is precisely the regime where FDL’s 32 TB/s bandwidth shines and where its lack of random-access capability barely registers. The notorious “memory wall” — the phenomenon where GPU clusters sit idle because memory bandwidth cannot keep pace with compute throughput — is exactly the problem this technology could help solve.

For Windows, Android, or any general-purpose operating system? The millisecond latency and sequential-only access would make the machine feel slower than a computer from 1995.

The Realistic Future

The most probable outcome is not disruption but division of labour. Traditional DRAM continues to handle the random-access general computing that defines everyday use. Fibre delay-line caching handles ultra-high-bandwidth sequential streaming in AI supercomputing clusters, operating as a specialised tier in a heterogeneous memory hierarchy — not as DDR5’s replacement, but as its complement in a domain where it has no business competing anyway.

Carmack’s idea is legitimate, the mathematics are sound, and the concept deserves serious research attention. What it does not deserve is the breathless “DDR5 is dead” framing that has dominated the popular coverage — a framing that confuses a precision instrument for a general-purpose tool, and in doing so, undersells what is genuinely a clever and potentially impactful concept.