Sorting the frontier

Sort the frontier by how real each part is

Shipping todaypost-quantum key exchangespecialised AI chipsheadset runtimesScaling nowhyperscale trainingthe energy billneuromorphic pilotsEmergingfault-tolerant quantumAI capability scalingphotonic computingSpeculativeDNA storagemolecular computingreversible computingPermanentspeed of lightheat per mm² of chipenergy to erase a bitSort the frontier by how real each part is: shipping, scaling, emerging, speculative, and limits that never move.Three shifts touch every engineer: the maths under encryption is being replaced, chips are specialising for AI, and siliconitself is being questioned for the first time since the transistor won.

The frontier runs from things already shipping to limits that physics will never move.

Sorting the frontier · Going deeper

The Frontier — what is shipping, what is research, and where the walls are Shipping today post-quantum crypto · specialised silicon · spatial computing runtimes Scaling now hyperscale training · the energy frontier · neuromorphic pilots Emerging fault-tolerant quantum · AGI capability scaling · photonic compute Speculative DNA storage · molecular compute · reversible computing The walls Landauer kT ln 2 per erase memory wall CPU vs DRAM gap speed of light 1 ft/ns floor cooling W per mm² the field has runway against most of these — and a few it does not

Shipping, research, and the walls

Three things are happening at once in 2026: the mathematics under TLS is being replaced; inference silicon is fracturing into specialised shapes because one architecture cannot win every workload within the power the grid can supply; and photonic, in-memory and DNA substrates are no longer only paper ideas.

Post-quantum cryptography

Recorded today, decrypted later: why encryption is changing now

todayencrypted traffic is recordedsome year: a large quantumcomputer existsa health record must stay secret this long: at risk nowa login that expires next week: finekey share a browser sends, bytes to scaleclassical X25519: 32post-quantum ML-KEM-768: 1,184classical secret32 bytespost-quantum secret32 bytesboth mixed into one session keyTraffic recorded today can be decrypted once a quantum computer exists, so key exchange is changing now.Browsers already send a hybrid: a classical and a post-quantum secret mixed into one key. Break one and the session holds.The standards, ML-KEM, ML-DSA and SLH-DSA, were finalised in August 2024. The price is size: about 2.3 KB more per handshake.

Post-quantum cryptography replaces the key-exchange maths that a quantum computer would break.

Post-quantum cryptography · Going deeper

A hybrid TLS 1.3 handshake with X25519 plus ML-KEM-768 key shares Client Server ClientHello key_share: x25519 (32 B) + ML-KEM-768 pk (1184 B) supported_groups: X25519MLKEM768, x25519, secp256r1 ServerHello key_share: x25519 (32 B) + ML-KEM-768 ct (1088 B) selected_group: X25519MLKEM768 derive derive ss_ec = X25519(a, B) ss_pq = ML-KEM.Decap(sk, ct) ss = HKDF(ss_ec ‖ ss_pq) ss_ec = X25519(b, A) ss_pq = ML-KEM.Encap → ct ss = HKDF(ss_ec ‖ ss_pq) Combined session secret secure if EITHER X25519 OR ML-KEM holds total handshake bytes grow by approximately 2.2 KB versus pure X25519

A hybrid TLS 1.3 handshake

The group X25519MLKEM768 carries a 32-byte X25519 key plus a 1,184-byte ML-KEM key (1,216 bytes); the server replies with 1,120. Both 32-byte secrets are concatenated into the TLS key schedule. Chrome enabled it by default in 2024. Migrate key exchange first, then code-signing signatures (ML-DSA signatures are 2.4 to 4.6 KB), then long-lived certificates.

Check yourself

2026recorded2038machine exists2045privacy needed until

An attacker records an encrypted transfer of medical files in 2026. Suppose a machine that breaks the key exchange appears in 2038, and the files must stay private until 2045. Are they safe?

How quantum computing works

Quantum speed comes from making wrong answers cancel

1 Start: every answer equally likely00011011measure now: random2 Mark: flip the sign of the right one00011011meanmeasure now: still random3 Reflect every bar about the mean00011011measure now: always 11A quantum algorithm arranges interference so the wrong answers cancel and the right one adds up.Grover’s search finds 1 item in N in about √N steps: here one step, not up to four. It halves the strength of symmetric keys,so AES-256 stays safe. Shor’s algorithm is the larger threat: it breaks RSA and elliptic curves outright.

A qubit register holds every answer at once with an amplitude each; algorithms steer the amplitudes.

How quantum computing works · Going deeper · 1 of 2

Grover's algorithm — amplitude amplification on a 4-item search Initial superposition After oracle After diffusion |00⟩ |01⟩ |10⟩ |11⟩ ½ ½ ½ ½ |00⟩ |01⟩ |10⟩ |11⟩ ← ½ ½ ½ −½ |00⟩ |01⟩ |10⟩ |11⟩ 0 0 0 1 one Grover iteration finds the marked item with certainty on N=4 — the rare exact case

Grover’s search, step by step

The mean of (½, ½, ½, −½) is ¼; reflecting a about it gives ½ − a. For N items the optimal count is about (π/4)√N iterations: a quadratic speed-up. AES-128 falls to roughly 64-bit strength, which is why AES-256 is recommended for long-lived data. Shor’s algorithm factors n-bit integers in about n³ operations: an exponential gap.

How quantum computing works · Going deeper · 2 of 2

A Bloch sphere with a tilted qubit state vector |0⟩ |1⟩ |−⟩ |+⟩ |ψ⟩ = α|0⟩ + β|1⟩ |α|² + |β|² = 1 θ φ α = cos(θ/2) β = e^(iφ) · sin(θ/2) measurement → |0⟩ with prob |α|² → |1⟩ with prob |β|² one qubit is a point on the surface; n qubits live in a 2ⁿ-dimensional complex space

The qubit

A qubit’s state is α|0⟩ + β|1⟩ with |α|² + |β|² = 1, drawn as a point on a sphere. Measurement gives 0 with probability |α|². n qubits carry 2ⁿ amplitudes, but one measurement returns one n-bit string. Entanglement, joint states that do not factor, is the resource algorithms consume.

Quantum code-breaking distance

A code-breaking quantum computer is about 10,000 times away

11,0001 million100 millionreliable (error-corrected) qubits demonstrated: single digits to low tensphysical qubits in the largest machines: a few thousandphysical qubits needed to break RSA-2048: about 20 millionabout 10,000 times more than existsThe machine that breaks today’s encryption needs about 10,000 times more qubits than any that exists.Qubits are noisy: about 1,000 physical ones are spent to make one reliable qubit. Nobody knows the date it arrives. Themigration is planned anyway, because recorded traffic cannot be re-encrypted afterwards.

Today’s machines are noisy; error correction spends about 1,000 physical qubits per reliable one.

Quantum code-breaking distance · Going deeper

NISQ vs fault-tolerant — the gap in physical qubits per logical qubit NISQ (2026) Fault-tolerant noisy intermediate scale error-corrected logical qubits physical qubits 10² – 10³ gate error 10⁻³ to 10⁻⁴ circuit depth hundreds before decoherence physical qubits 10⁶ – 10⁸ logical error 10⁻¹² target for Shor surface code overhead ≈1000 physical per logical a few demonstrated logical qubits in 2024-2025 — millions required for cryptographically relevant Shor

Noisy today, fault-tolerant later

Today’s NISQ machines: 10² to 10³ physical qubits, gate error 10⁻³ to 10⁻⁴, a few hundred gates before noise wins. Fault-tolerant machines need 10⁶ to 10⁸ physical qubits and a logical error near 10⁻¹². In 2024 and 2025 several groups showed logical qubits with lower error than their parts. No NISQ machine yet beats classical computers on a commercially relevant problem.

AI accelerators

AI chips trade flexibility for energy, ten times at a step

CPU server1 JGPU0.1 JTPU: matrix maths only30 mJPhone NPU10 mJCustom LPU5 mJAnalog (estimate)0.1 mJenergy to classify one image (ResNet-50); each gridline is 10 times less; order-of-magnitude figuresA chip that gives up flexibility for one fixed job cuts the energy per answer by up to ten times.GPUs won training because training is messy. Serving runs one model shape billions of times, so chips that only multiplymatrices win there. The risk: when model architecture changes, a chip built for the old shape is scrap.

Specialised silicon: chips built for one dataflow, such as matrix multiplication, and nothing else.

AI accelerators · Going deeper · 1 of 2

A 4×4 systolic array — weights stay, activations flow, partial sums accumulate MAC MAC MAC MAC MAC MAC MAC MAC MAC MAC MAC MAC w₀₀ w₀₁ w₀₂ w₀₃ w₁₀ w₁₁ w₁₂ w₁₃ w₂₀ w₂₁ w₂₂ w₂₃ a₀ → a₁ → a₂ → ↓ partial sums Systolic array — activations flow right, partial sums flow down no shared bus, no register file, no scheduling — just a grid that does matmul

A systolic array: the data movement is the operation

Google’s first TPU (deployed 2015) was a 256×256 grid of multiply-accumulate cells: activations flow one way, weights the other, sums accumulate. Groq’s LPU schedules every instruction at compile time, with no caches or branch predictors: 500 to 700 tokens a second. Cerebras uses a whole wafer as one chip: 900,000 cores and 44 GB of on-chip memory.

AI accelerators · Going deeper · 2 of 2

Per-inference energy across silicon families for a ResNet-50 forward pass 10⁰ 10⁻¹ 10⁻² 10⁻³ 10⁻⁴ joules per inference CPU GPU TPU NPU LPU neuro ≈1 J ≈0.1 J ≈30 mJ ≈10 mJ ≈5 mJ est. ≈0.1 mJ ResNet-50 inference — energy per image, log scale

Energy per inference across chip families

Rough order-of-magnitude figures from MLPerf results and vendor disclosures, not benchmarks. A phone NPU runs 30 to 50 trillion operations a second on under 5 W. FPGAs are the reconfigurable middle: about 10 times slower and less dense than a fixed chip, but no fabrication cycle for a new model.

Data movement and energy

Moving a number costs far more than computing with it

Memoryweights live hereComputemultiply herethe busfetch one weight from memory: 640 pJmultiply with it: 1 pJbars to scaleMemory and compute in one placeneuromorphic chips, in-memory computingspikes: only the neurons that fire use energythe brain: 86 billion neurons on about 20 WIn a conventional chip, moving a number costs hundreds of times more energy than computing with it.Neuromorphic and in-memory designs keep the weights where the arithmetic happens and work only when something changes.Shipping today in always-on sensors and wake-word chips; still research for general deep learning.

The von Neumann bottleneck: memory and compute sit apart, and the trip between them dominates.

Data movement and energy · Going deeper

Spiking-neuron timing — membrane potential, threshold, and output spikes V t θ spike spike integrate inputs · leak slowly spike when V crosses θ · reset Leaky integrate-and-fire neuron

Spiking neurons and in-memory computing

A spiking neuron sums incoming spikes, leaks toward rest, and fires on crossing a threshold. Intel’s Loihi 2 has a million neurons per chip; IBM’s TrueNorth ran a million neurons on 70 mW. A crossbar of resistive cells multiplies a vector by a matrix in one physical step, at 1 to 10 fJ per operation in the lab. The open problem is training: accuracy still trails ordinary digital networks.

Spatial computing latency

A headset has 11 ms, so it predicts your head, then corrects

a screen on a desk: 30 ms of delay goes unnoticeda screen on your head: 11 ms from head movement to light, or people feel sickSample pose1,000 times a secondPredict10 to 15 ms aheadRenderthe app draws the frameWarpshift to the latest poseDisplaypredict: aim the picture at where the head will be, not where it iswarp: just before display, read the head position again and shift the finished picture to matchA headset has 11 ms from head movement to light. It gets there by predicting the head, then correcting.Eye tracking adds foveated rendering: full detail only where you are looking, for 40 to 60% less drawing work.Whether headsets become an everyday computer or stay a niche is still open; the runtime underneath works.

Motion-to-photon latency: the time from a head movement to the light that shows it.

Spatial computing latency · Going deeper

Motion-to-photon timing pipeline for a VR frame 0 ms 3 ms 6 ms 8 ms 10 ms 11 ms IMU predict render timewarp scan-out sample pose extrapolate draw frame warp to new pose display photons 11 ms wall From head movement to photon — the entire budget miss the wall and the user feels lag; miss it badly and they feel sick

The motion-to-photon pipeline

OpenXR is the cross-vendor API. Asynchronous timewarp re-projects the last rendered frame to the newest pose; spacewarp synthesises a frame the application dropped. With eye tracking, foveated rendering shades only the central 2 to 3 degrees at full rate. It is the closest a consumer product comes to hard real-time scheduling.

Computing and electricity

Computing is now an electricity problem

of every watt drawn from the gridolder data centre (PUE 2.0)half reaches the chipsmodern hyperscale (PUE 1.1)91% reaches the chipspower in one rack, to scaleair cooling tops out: 30 kWan AI rack: 80 to 120 kW, liquid-cooledelectricity to train one frontier model, to scale2020: 1,287 MWh2024 to 2025: 50,000 to 200,000 MWh. Tens of megawatts, for weeksComputing has become an electricity problem: one training run draws tens of megawatts for weeks.Data centres used about 460 TWh in 2022 and are on course to double by 2026, mostly from AI. The same job can emit halfthe carbon if it runs when the grid is on sun and wind: carbon-aware scheduling.

PUE, power usage effectiveness: total power drawn divided by the power that reaches the computers.

Computing and electricity · Going deeper

Data-center cooling — heat flow from die to atmosphere Die Cold plate CDU Cooling tower 100 °C 65 °C 45 °C ambient Immersion variant — server submerged in dielectric fluid no cold plates · fluid carries heat directly · density ≈ 100 kW per rack Air cooling — fans, hot/cold aisles, CRAC units legacy default · density tops out at ≈ 30 kW per rack · approaches limit for H100/B200 racks From transistor heat to atmosphere

Cooling: from the chip to the atmosphere

Direct-to-chip liquid cooling puts cold plates on each chip; immersion submerges whole servers. Either roughly halves cooling’s share. A long LLM response costs roughly 1 to 10 watt-hours. Operators pair sites with renewables, geothermal and nuclear. A kilowatt-hour budget and a carbon target are becoming part of the engineer’s job.

Training at scale

At 100,000 chips, failure is routine and returns diminish

one training job on 10,000 GPUs: between 0.5 and 2 of them fail every daya checkpoint every hour0 h8 ha GPU failsonly this work is redonewhat the hardware could compute at full speed35 to 50% does useful arithmeticthe rest waits on the network and on other chipsAt 100,000 chips, failure is routine, and each added chip gives less than the one before.On every step all the chips must exchange results, so the network between them matters as much as the chips. The part ofa job that cannot be split caps the speed-up however much hardware is added: Amdahl’s law.

Hyperscale training: one job across 10⁵ chips, built around checkpoints and communication.

Training at scale · Going deeper · 1 of 2

3D parallelism — data, tensor, and pipeline dimensions on the same cluster Data parallel Tensor parallel Pipeline parallel replicate model · split batch split layer · within node split layers · across nodes model model copy copy all-reduce gradients W₀ W₁ W₂ W₃ all-gather activations layers 1-8 layers 9-16 layers 17-24 layers 25-32 micro-batch pipeline a 24,576-GPU run might be 8-way tensor × 96-way pipeline × 32-way data

Three ways to split one model

Data parallelism copies the model and splits the batch. Tensor parallelism splits one layer’s matrices across chips in a node. Pipeline parallelism puts different layers on different chips. Real runs combine all three. A checkpoint of a 100-billion-parameter model is 1 to 2 TB and must be written in under a minute.

Training at scale · Going deeper · 2 of 2

Ring all-reduce across N GPUs — bandwidth-optimal gradient sync G₀ G₁ G₂ G₃ G₄ G₅ step 1: scatter-reduce each node owns a shard step 2: all-gather broadcast shard around ring 2(N−1) hops total bandwidth ≈ 2·data/N independent of N at the limit the ring is bandwidth-optimal for large messages — every link runs at line rate the whole time

Ring all-reduce

Every chip contributes a gradient and receives the sum. The ring algorithm splits the tensor into N shards and passes them around in 2(N−1) hops, so bandwidth per chip stays constant as N grows. NVLink carries 900 GB/s inside a node; between nodes, RDMA writes straight into remote memory in microseconds.

Check yourself

95%: can run in parallel5%: one at a time

A job is 95% work that can be split across chips and 5% that cannot. With unlimited chips, what is the greatest possible speed-up?

Scaling laws

More compute has reliably meant more capability. So far

error (lower is better)compute, model size and data: each step to the right is 10 times more (schematic)GPT-2, 2019GPT-4, 20232025 to 2026the same slope?smaller returns?A newer levermore compute per answer: reasoningSo far each ten-fold increase in compute has bought a predictable gain. Nobody knows whether that continues.Old benchmarks are saturated, so testing frontier models is now its own discipline. What has converged is practice: capabilitytests before release, staged access, monitoring, and regulation such as the EU AI Act.

Scaling laws: model error falls as a smooth power law in compute, parameters and data.

Scaling laws · Going deeper

Scaling laws — loss as a power law in compute, parameters, and tokens 2.0 3.0 4.0 5.0 test loss 10¹⁹ 10²¹ 10²³ 10²⁵ 10²⁷ training compute (FLOPs, log scale) GPT-2 GPT-3 PaLM / Chinchilla GPT-4 era 2025-26 frontier ? Loss falls as a power law — until it might not

What is contested and what has converged

There is no agreed definition of “general” intelligence. Benchmarks such as MMLU saturated and were replaced by harder ones such as SWE-bench Verified. Alignment, making behaviour match intent as capability grows, feeds practice through red-teaming, capability evaluations tied to scaling policies, and post-deployment monitoring.

Physical limits

Some limits are physics. Others are only expensive

Speed of lightNew York to London: 28 ms one way, at the very bestphysics, and felt todayEnergy to erase a bit2.85 × 10⁻²¹ J; chips use about a million times morephysics, still far offHeatabout 1 W per mm² of chip before cooling gets exoticcost: bends with moneyMemory300 to 500 instructions pass while one memory fetch returnscost: bends with moneySerial work5% that cannot be split caps the speed-up at 20 timesthe algorithm: change itSome limits are physics and some are only expensive. Knowing which is which tells you what can change.Light speed is why computing moves to the edge, close to users. The candidates beyond silicon, photonics for matrix mathsand DNA for archives, each attack one wall. None replaces digital logic yet.

The walls: light speed and the Landauer limit are absolute; heat, memory and cost bend at a price.

Physical limits · Going deeper · 1 of 2

The walls — log-scale ceilings on energy, latency, and bandwidth 10⁻¹² 10⁻¹⁵ 10⁻¹⁸ 10⁻²¹ J / bit Landauer 1990 2000 2010 2020 2030? transistor switching energy versus the Landauer floor CMOS switching energy approaching the thermodynamic floor six orders of magnitude of headroom remain — most will be spent on the way down

Switching energy against the thermodynamic floor

Landauer (1961): erasing a bit costs at least kT ln 2, about 2.85 × 10⁻²¹ J at room temperature. Transistors switch at around 10⁻¹⁵ J. Reversible computing (Bennett, 1973) is the only theoretical way under the floor and is not practical. A geostationary satellite adds 120 ms each way; low Earth orbit, 2 to 3 ms.

Physical limits · Going deeper · 2 of 2

The Frontier — what is shipping, what is research, and where the walls are Shipping today post-quantum crypto · specialised silicon · spatial computing runtimes Scaling now hyperscale training · the energy frontier · neuromorphic pilots Emerging fault-tolerant quantum · AGI capability scaling · photonic compute Speculative DNA storage · molecular compute · reversible computing The walls Landauer kT ln 2 per erase memory wall CPU vs DRAM gap speed of light 1 ft/ns floor cooling W per mm² the field has runway against most of these — and a few it does not

Beyond silicon

Transistor features are at 2 to 3 nm; clock speeds have barely risen since about 2005. Photonic chips multiply matrices with light at a claimed 10 to 100 times less energy, and ship for narrow jobs. DNA could hold about 10¹⁸ bytes per gram and last thousands of years, but writes at hundreds of bases a second: archive only. Molecular computing is academic. Treat most of this as research.

Check yourself

Asignal delay,New York to LondonBthe wait formemory fetchesCminimum energyto erase a bit

A company with unlimited money wants to push back one of these limits. Which one can its engineers actually move?

Summary

What each layer is made of changes. The map stays

Interfaces: peopleSystems, data and AINetworks and securityMachines and chipsBitsWhat staysthe layers and their contractsthe trade-offs: speed, cost, safetythe limits: light, heat, energyWhat changeswhat each layer is built fromheadsets on an 11 ms budgetcapability still scaling; energy is the billpost-quantum keys, shipping nowspecialised now, perhaps photonic nextstill the unit, whatever holds itThe frontier changes what each layer is built from. The layers, the trade-offs and the limits stay.That is the map: sixteen lessons, from a single bit to the edge of what can be built. Every other course here goes deeperinto one part of it.

You now have the whole map: from one bit to the edge of what can be built.

Lesson complete

You finished “Frontier”

Answer the questions along the way to see your score here.

Back to Digital Systems: the Map

Open in a new tab