LineShine: When Raw Speed Breaks Encryption

$NVDA · $AMD · $INTC · $ARM · $AVGO · $MRVL · $MU · $HPE · $DELL · $SMCI · $IBM · $MSFT · $GOOGL · $TSM

In June 2026, the world's most powerful computer stopped being American — and, just as importantly, it stopped depending on a single Nvidia or AMD graphics chip to get there. A machine called LineShine, installed at the National Supercomputing Centre in Shenzhen, topped the 65th edition of the TOP500 list with a sustained High-Performance Linpack score of 2.198 exaflops — more than two quintillion double-precision calculations every second. It is the first system in the history of the ranking to break the two-exaflop barrier on that benchmark, and it did it with nothing but CPUs.

That single fact reorganises a lot of assumptions. For three years, the conversation about computing's threat to encryption has been almost entirely about quantum machines — exotic, cryogenically cooled, error-prone devices that, in theory, can unwind the mathematics protecting the world's data. Quantum is real and it is coming. But LineShine is a reminder that the danger to legacy systems is not the type of computer. It is the speed. Raw, brute, conventional processing power is growing faster than most security planners have priced in, and every doubling shortens the runway to the moment our current cryptography stops being safe — what the industry calls Q-Day.

This is the case that matters for anyone running a bank, an exchange, a hospital, a government archive, or any system that assumes today's encrypted data will still be private in ten years. Whether the blow comes from a quantum processor or from a hall full of fast CPUs, the outcome is the same: data you thought was locked becomes readable. Let's look at what LineShine actually is, how it stacks up against quantum on both speed and cost, why its arrival tightens the clock on Q-Day, and which publicly listed companies — in mainland China, Hong Kong, and the United States — are building the machines on both sides of this race.

What LineShine actually is

LineShine is not a science-fiction object. It is a very large, very well-engineered conventional supercomputer, and its specifications tell the story of a deliberate strategy.

The system is built from 40,960 LX2 processors spread across 92 cabinets. Each LX2 chip carries 304 cores based on the Armv9 instruction set, clocked at 1.55GHz. Crucially, every core includes Arm's Scalable Vector Extension (SVE) and Scalable Matrix Extension (SME) units — the hardware that accelerates the vector and matrix math at the heart of both scientific simulation and AI training. Add it up and LineShine fields more than 13 million CPU cores, all of them general-purpose, none of them a dedicated GPU accelerator of the kind that powers every other machine at the top of the list.

The cores are stitched together by a proprietary interconnect called LingQi, using a dual-plane, multi-rail fat-tree topology delivering 1.6 terabits per second to every node. Memory sits close to compute: each LX2 socket holds two compute dies divided into four NUMA domains, with high-bandwidth memory attached to each domain and a dedicated data-movement engine to keep the cores fed. Storage runs across 428 nodes with roughly 10 terabits per second of aggregate bandwidth. The software stack is domestic too — the machine runs Kylin OS on Huawei Kunpeng server infrastructure.

On the standard benchmarks, LineShine didn't just win the headline race. It also took the No. 1 spot on the HPCG benchmark — which measures performance on the messy, memory-bound workloads that real science actually runs — at 22.0 petaflops, and placed fourth on the mixed-precision HPL-MxP test at 7.92 exaflops. In plain terms, this is not a one-trick benchmark special. It is a broadly capable machine.

What it dethroned matters. The previous champion, the United States' El Capitan at Lawrence Livermore National Laboratory, holds roughly 1.8 exaflops. Oak Ridge's Frontier, the first machine ever to break the exaflop barrier back in 2022, now sits third at about 1.35 exaflops. LineShine is the first Chinese system to lead the global ranking since 2017, and it arrived after a multi-year stretch in which China declined to submit its fastest machines to the list at all.

The all-CPU gambit

The most strategically interesting thing about LineShine is what it does not contain: any Nvidia H100, B200, or AMD Instinct GPU. Since 2022, U.S. export controls have progressively cut China off from the most advanced AI accelerators and the equipment to make them at scale. The conventional wisdom held that this would freeze China out of the leading edge of high-performance computing.

LineShine is the counter-argument. By leaning on huge fleets of domestically designed Arm-based CPUs — with vector and matrix extensions doing the heavy lifting that GPUs normally handle — the designers built a top-of-the-world machine without a single sanctioned part. It is, above all, a statement about compute sovereignty: the ability to reach the frontier of processing power using only hardware and software you fully control.

There is a cost to this design, and it shows up in the power bill. LineShine draws 42.2 megawatts at full tilt — the highest power consumption of any machine in the global top ten — and lands only 50th on the Green500 efficiency ranking at 52.07 gigaflops per watt. El Capitan, by contrast, delivers more efficient performance per watt because GPU accelerators are simply better at dense math per joule. So the all-CPU approach trades electrical efficiency for supply-chain independence. For a state that has decided self-reliance is non-negotiable, that is a trade worth making — and it tells you the appetite for raw compute is now strong enough to absorb the inefficiency.

Exascale in four years

Step back from the single machine and look at the curve, because the curve is the real story.

In 2016, the fastest computer on Earth was China's Sunway TaihuLight, at 93 petaflops. In 2018 the U.S. retook the lead with Summit at 149 petaflops. Japan's Fugaku reached 442 petaflops in 2020. Then Frontier crossed the exaflop line in 2022 at 1,102 petaflops — a number that had been treated as a distant, almost mythical milestone for over a decade. El Capitan hit roughly 1,742 petaflops in late 2024. And now LineShine sits at 2,198 petaflops in 2026.

Read those numbers as a trend and the implication is stark: the performance of the world's single fastest machine has more than doubled in four years, and it has roughly 24-folded in a decade. This is faster than the old Moore's Law cadence of doubling every two years that defined the industry for half a century — because the gains now come not only from denser transistors but from massive parallelism, specialised math units, faster interconnects, and the sheer willingness of states to pour gigawatts and billions of dollars into the problem.

And the frontier machines are only the visible tip. Below them sits an exploding population of commercial AI superclusters — the GPU and accelerator farms built by hyperscalers to train large language models — many of which would rank near the top of the list if their owners bothered to submit them. The total amount of computing power available to a well-funded adversary, whether a nation-state intelligence agency or a sophisticated criminal group, is rising on a steeper slope than the official rankings alone suggest. The next named target after exascale — zettascale, a thousand times faster than Frontier — is already on national roadmaps for the early 2030s.

Two roads to the same cliff

Here is where the encryption question comes in, and where it pays to be precise, because quantum computing and classical supercomputing threaten cryptography in genuinely different ways.

The quantum road is structural. The encryption that secures almost everything online — RSA, Diffie-Hellman, and elliptic-curve cryptography — rests on math problems (factoring large numbers, computing discrete logarithms) that are effectively impossible for ordinary computers to solve at scale. A sufficiently large, error-corrected quantum computer running Shor's algorithm doesn't just do that math faster; it changes the complexity class of the problem, collapsing a calculation that would take a classical machine longer than the age of the universe into one that finishes in hours or days. Quantum is a skeleton key for a specific, dominant family of cryptography.

The classical road is brute force and cleverness. A conventional supercomputer cannot magic away the math the way Shor's algorithm does. But raw speed still erodes security in several concrete ways:

• It directly shrinks the safety margin of symmetric ciphers and hash functions. Algorithms once considered comfortably out of reach become brute-forceable as the number of operations per second climbs. Short RSA keys, legacy ciphers, weak hashes, and poorly chosen parameters that were "good enough" a decade ago fall to a machine doing two quintillion operations a second.

• It supercharges cryptanalysis — the search for mathematical shortcuts and implementation flaws. Modern attacks rarely brute-force a whole keyspace; they exploit weaknesses, and finding those weaknesses is itself a massive search-and-simulation problem that scales directly with available compute.

• It accelerates the research that builds the quantum threat. Designing better qubits, simulating error-correcting codes, and optimising quantum algorithms are all done on classical supercomputers. The faster the classical machines, the faster the quantum machines arrive. The two roads are not parallel — they feed each other.

• Combined with AI, fast compute changes the economics of attack. Machine-learning models trained on enormous clusters are increasingly used to find side-channel leaks, predict weak random-number generation, and accelerate lattice-reduction and other cryptanalytic techniques. The AI is only as strong as the compute behind it, and the compute is growing fast.

The point the security world has been slow to absorb is this: you do not need a quantum computer to be hurt by the speed curve. The legacy systems most exposed to Q-Day — the ones running outdated cipher suites, hard-coded keys, and decades-old protocols — are also the ones most exposed to a determined adversary with a few tens of megawatts of conventional compute today. Speed is the danger, whatever its source.

How LineShine tightens the clock on Q-Day

Q-Day is shorthand for the moment a cryptographically relevant quantum computer first exists — the day RSA-2048 and its cousins can be broken on demand. Nobody knows the exact date. What we do know is that the estimate keeps moving closer, and faster classical computing is one of the reasons.

Consider how the published estimates for breaking RSA-2048 have collapsed. In 2012, researchers pegged the requirement at around a billion physical qubits. In 2019, Craig Gidney and Martin Ekerå brought that down to roughly 20 million physical qubits running for about eight hours. Then in 2025, Gidney — now at Google Quantum AI — published a new analysis arguing the job could be done with under one million qubits in under a week. That is a twenty-fold reduction in the hardware bar in six years, driven almost entirely by better algorithms and better error-correction schemes — work performed on classical computers. Every gain in classical simulation capacity lets researchers test more aggressive designs, which is precisely how the qubit requirement keeps falling.

Now layer on the threat that is already live, no quantum computer required: harvest now, decrypt later (HNDL). Adversaries are recording encrypted traffic and exfiltrating encrypted archives today, betting they can decrypt them once the capability — quantum or classical — matures. The clock on HNDL does not start on Q-Day; it started the moment the data was captured. Anything with a long secrecy lifetime — health records, state secrets, financial ledgers, biometric data, intellectual property, the private keys behind digital assets — is already at risk if it is sitting in someone's cold storage waiting for the speed curve to catch up. A machine like LineShine, and the dozens of commercial superclusters being built alongside it, lowers the cost of that eventual decryption and brings the payoff date forward.

There is a subtler effect too. As bulk compute gets cheaper, the threshold of who can afford an attack drops. Capabilities once reserved for two or three superpowers become available to mid-tier states and well-funded criminal enterprises. Q-Day, in practice, is not a single global event — it is a staggered series of moments, each one arriving when a given class of attacker can finally afford the compute to break a given class of target. LineShine is evidence that the cost of frontier compute is falling even under sanctions. That should worry anyone whose threat model assumed the attacker would be resource-constrained.

The cost of speed

Cost is where the quantum-versus-classical comparison gets genuinely interesting, and where the asymmetry between attacker and defender becomes clear.

El Capitan, the machine LineShine just overtook, cost a reported $600 million to build and consumes more than 35 megawatts — roughly the power draw of a small town. LineShine's price has not been disclosed, but its 42.2-megawatt appetite and 40,960 custom processors put it firmly in the same nine-or-ten-figure class. These are nation-state instruments. But here is the thing: the per-unit cost of computation on these machines keeps falling. Each generation delivers far more work per dollar and per watt than the last, even as the total system price stays high. That is the economic engine driving the speed curve, and it applies equally to the commercial AI clusters that any cash-rich corporation — or criminal syndicate laundering serious money — can now rent by the hour.

Quantum carries a different and, for now, far steeper cost structure. A cryptographically relevant quantum computer requires not just a million-plus physical qubits but the cryogenic plumbing to hold them near absolute zero, the control electronics to manipulate them, and the staggering overhead of quantum error correction — where thousands of noisy physical qubits are lashed together to produce a single reliable logical qubit. One analysis of the energy required to actually run such a factoring attack puts it in a range that is anything but trivial. Quantum is not a cheap shortcut; it is an enormously expensive megaproject that only the best-funded players on Earth can attempt — today.

The strategic read is this: classical speed is the near-term, broadly affordable threat; quantum is the medium-term, structural one. Defenders have to plan for both at once, and the cost trajectory of classical compute means the near-term threat is escalating right now, on a curve LineShine just made visible.

Quantum's parallel clock

None of this means quantum is far off or unimportant — the opposite. The headline machines are advancing on their own steep curve. IBM put a 1,121-qubit processor into the field in 2023 and has a public roadmap toward large fault-tolerant systems by the end of the decade. Google's Willow chip, unveiled in late 2024, demonstrated that adding more physical qubits can reduce error rates rather than increase them — the long-sought signal that error correction scales in the right direction. Neutral-atom and trapped-ion approaches from a clutch of companies are pushing qubit counts and fidelities from other angles entirely.

The honest summary is that no one has broken RSA-2048, and no machine in existence today is close. But the direction of travel is unambiguous, the qubit bar is falling, and the classical machines doing the design work are getting faster every year. The two clocks — classical and quantum — are running together, and they point at the same destination.

The defensive answer is already standardised

The reassuring part of this story is that the fix exists and is no longer experimental. In August 2024, the U.S. National Institute of Standards and Technology finalised the first three post-quantum cryptography standards: FIPS 203 (ML-KEM, for key exchange — the replacement for RSA and Diffie-Hellman key agreement), FIPS 204 (ML-DSA, for digital signatures), and FIPS 205 (SLH-DSA, a hash-based signature backup). These algorithms are designed to resist both quantum and classical attack, and they run on the ordinary hardware organisations already own.

The deadlines are now concrete. The U.S. National Security Agency's CNSA 2.0 framework calls for quantum-resistant algorithms in new national-security systems from 2027, broad software migration by 2030, and full infrastructure migration by 2035 — with RSA-2048 and the P-256 elliptic curve formally deprecated along the way. For most institutions the binding question is no longer whether to migrate but how fast, and the harvest-now-decrypt-later threat means data with a long shelf life should be re-protected first, not last.

This is the discipline KXCO has built around: signing every new piece of published work and securing session and identity infrastructure with NIST-standardised, Level-3 post-quantum algorithms (ML-DSA-65 and ML-KEM-768) rather than waiting for a breach to force the issue. The principle is simple and it applies to any serious operator: the migration takes years, the threat is already accumulating, and the only sensible time to start was yesterday.

What a serious operator does now

The temptation, faced with a threat this large and this uncertain in timing, is to wait for more clarity. That is the wrong instinct, because the migration itself is slow and the threat compounds quietly. A credible programme has a handful of moving parts, and none of them require a quantum computer to justify starting today.

The first is discovery. Most large organisations genuinely do not know where their cryptography lives — which systems use which algorithms, which keys protect which data, which third-party services and embedded devices are quietly relying on RSA or elliptic-curve under the hood. You cannot migrate what you cannot see, so building a cryptographic inventory is the unglamorous but essential first step.

The second is triage by data lifetime. Not all data is equally exposed to harvest-now-decrypt-later. A session token that expires in an hour barely matters; a patient record, a sealed legal file, a custody key, or a state secret that must stay private for twenty years is a priority-one target the moment it crosses a network. Re-protect long-lived secrets first, working backward from how long the data must stay confidential.

The third is crypto-agility — designing systems so the algorithm can be swapped without rebuilding the whole application. Many of the legacy systems most at risk are precisely the ones where cryptography was hard-coded a decade or two ago and cannot be changed without a major project. The institutions that invest in agility now will migrate again cheaply when the standards evolve, as they inevitably will.

The fourth is hybrid deployment. The current best practice is not to rip out classical cryptography overnight but to run post-quantum algorithms alongside the existing ones, so a flaw newly discovered in either scheme still leaves the other protecting the data. This is the conservative path the major browsers and cloud providers have already taken for key exchange, and it is available to any operator today.

The point that ties back to LineShine is that every one of these steps is justified by the classical speed curve alone, before quantum even enters the picture. Faster brute force, cheaper bulk compute, and AI-assisted cryptanalysis are reasons enough to harden weak ciphers, retire short keys, and re-protect long-lived data now. Quantum simply raises the stakes and sets a hard deadline on work that was already overdue.

The investable map: who builds the machines

For investors and strategists trying to track this race through public markets, here are the listed companies most directly exposed — on both the speed side and the quantum side. (Note one large caveat at the centre of the LineShine story: Huawei, whose Kunpeng platform and chip designs underpin the machine, is privately held and not listed anywhere; much of China's most strategic compute capability sits outside the public markets entirely.)

Mainland China (Shanghai / Shenzhen listed)

• Cambricon Technologies (688256.SH) — China's leading designer of dedicated AI accelerator chips, often called the country's answer to Nvidia.

• Hygon Information Technology (688041.SH) — x86-compatible server CPUs and DCU accelerators for data centres and HPC.

• Loongson Technology (688047.SH) — developer of the home-grown LoongArch CPU architecture used in domestic computers and supercomputers.

• Dawning Information Industry / Sugon (603019.SH) — one of China's principal builders of supercomputers and HPC servers (placed on the U.S. Entity List).

• Inspur Electronic Information (000977.SZ) — a top global server manufacturer and a backbone supplier of HPC and AI hardware (also on the U.S. Entity List).

• SMIC — Semiconductor Manufacturing International (688981.SH) — China's largest chip foundry, the fabrication base for domestic processors.

• Hua Hong Semiconductor (688347.SH) — the second major domestic foundry.

• Montage Technology (688008.SH) — memory-interface and interconnect chips critical to server and HPC performance.

Hong Kong listed

• SMIC (0981.HK) and Hua Hong Semiconductor (1347.HK) — the two leading foundries, dual-listed in Hong Kong.

• Lenovo Group (0992.HK) — a global top-tier server and HPC systems builder.

• ZTE (0763.HK) — telecom and computing infrastructure, including server and chip work.

• Alibaba Group (9988.HK) — through its T-Head / PingTouGe unit, a designer of Arm-based server and AI chips, plus a hyperscale cloud.

• Baidu (9888.HK) — developer of the Kunlun line of AI accelerator chips.

• Tencent (0700.HK) — hyperscale cloud and in-house silicon investment.

• Kingsoft Cloud (3896.HK) — cloud infrastructure (also NASDAQ-listed as KC).

• Horizon Robotics (9660.HK) — AI compute chips, listed in Hong Kong in 2024.

United States listed

On the classical-compute side:

• NVIDIA (NVDA) — the dominant supplier of AI and HPC accelerators worldwide.

• AMD (AMD) — builder of the CPUs and Instinct GPUs inside El Capitan and Frontier.

• Intel (INTC) — CPUs and the Aurora exascale system's accelerators.

• Arm Holdings (ARM) — owner of the Armv9 architecture that LineShine itself is built on; it profits from Arm-based designs regardless of who ships them.

• Broadcom (AVGO) and Marvell (MRVL) — custom silicon and the high-speed networking that ties superclusters together.

• Micron Technology (MU) — high-bandwidth memory, the scarce ingredient in modern AI and HPC systems.

• Hewlett Packard Enterprise (HPE) — the Cray division built both El Capitan and Frontier.

• Dell Technologies (DELL) and Super Micro Computer (SMCI) — leading builders of AI and HPC servers.

• IBM (IBM) — supercomputing heritage and a leading quantum-hardware programme.

• Microsoft (MSFT) and Alphabet (GOOGL) — hyperscale clouds, in-house AI silicon, and major quantum research arms (Google Quantum AI built the Willow chip).

• Taiwan Semiconductor (TSM) — the foundry that fabricates most of the world's leading-edge compute, U.S.-listed via ADR.

On the quantum and quantum-security side:

• IonQ (IONQ) — trapped-ion quantum systems.

• Rigetti Computing (RGTI) — modular superconducting quantum processors.

• D-Wave Quantum (QBTS) — quantum annealing and gate-model systems.

• Quantum Computing Inc. (QUBT) — photonic quantum and related hardware.

• Arqit Quantum (ARQQ) — post-quantum encryption software and key distribution.

This is not a recommendation to buy or sell any of these names — several of the pure-play quantum stocks routinely swing 10–15% on no news and have suffered repeated 50%-plus drawdowns. It is a map of where the capability is being built and where the public-market exposure to this race actually sits.

The bottom line

LineShine is a landmark for several reasons — China's return to the top of the rankings, the first machine past two exaflops, the proof that a frontier supercomputer can be built without Western accelerators. But the lesson that should travel furthest has nothing to do with national bragging rights. It is that computing speed is now the dominant variable in the security of every legacy system on the planet, and that speed is rising faster, on more fronts, and for more players than the quantum-focused conversation has allowed.

Quantum will eventually deliver a structural break in our cryptography. Classical speed is already eroding it at the margins and getting cheaper by the year. Both clocks point at Q-Day, and machines like LineShine are turning the hands faster. The organisations that come through this intact will be the ones that stopped treating post-quantum migration as a future project and started treating it as present-tense maintenance — because the data being harvested today will be read on whatever machine gets fast enough first, and that machine is arriving sooner than the calendar suggests.

Mainland China (Shanghai/Shenzhen)
$688256 Cambricon · $688041 Hygon · $688047 Loongson · $603019 Sugon · $000977 Inspur · $688981 SMIC · $688347 Hua Hong · $688008 Montage

Hong Kong
$0981 SMIC · $1347 Hua Hong · $0992 Lenovo · $0763 ZTE · $9988 Alibaba · $9888 Baidu · $0700 Tencent · $3896 Kingsoft Cloud · $9660 Horizon Robotics

United States — classical compute
$NVDA · $AMD · $INTC · $ARM · $AVGO · $MRVL · $MU · $HPE · $DELL · $SMCI · $IBM · $MSFT · $GOOGL · $TSM

United States — quantum / quantum-security
$IONQ · $RGTI · $QBTS · $QUBT · $ARQQ

Sources: TOP500 — 65th list, June 2026; Tom's Hardware — "China's LineShine supercomputer dethrones El Capitan"; DataCenterDynamics — "LineShine: All-CPU Chinese supercomputer named world's most powerful"; Network World — "China's LineShine dethrones El Capitan"; Lawrence Livermore National Laboratory — El Capitan; HPCwire — "Exascale Is Officially Here with Debut of Frontier"; MIT Technology Review — breaking 2048-bit RSA; Google Quantum AI / Craig Gidney, 2025 RSA-factoring estimate; NIST — Post-Quantum Cryptography Standards FIPS 203/204/205 (August 2024); NSA — CNSA 2.0 migration timeline.