The frontier runs from things already shipping to limits that physics will never move.
This last lesson looks ahead. And the first skill when looking ahead is telling what is real from what is hoped for. Some of the frontier is already shipping. Your browser is using new encryption designed to survive quantum computers. Phones and data centres run chips built only for AI. Some is scaling right now: training runs across a hundred thousand chips, and the electricity bill that comes with them. Some is emerging: quantum computers that correct their own errors, and the open question of how far AI capability scales. Some is speculative: storing data in DNA. And some things are permanent. The speed of light. The heat a chip can shed. We will take them in that order, with the numbers.
Sorting the frontier · Going deeper
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
Post-quantum cryptography replaces the key-exchange maths that a quantum computer would break.
Start with what is already shipping, and why it could not wait. Today, an attacker can record encrypted traffic. They cannot read it. They can store it. In some future year, a large enough quantum computer exists, and it breaks the mathematics that today’s key exchange depends on. Everything recorded becomes readable. So ask how long a secret must last. A health record must stay private for decades: it is at risk now. A login that expires next week is not. The fix is new mathematics, and its cost is size. A classical key share is thirty-two bytes. The post-quantum one is one thousand one hundred and eighty-four. And nobody fully trusts the new mathematics yet. So browsers send both, and mix the two secrets into one key. To read the session, an attacker must break both.
Post-quantum cryptography · Going deeper
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
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?
Expiry does not help: the attacker holds the recorded exchange and derives the key from it. That is “harvest now, decrypt later”, and why long-lived data needs post-quantum key exchange today.
How quantum computing works
Quantum speed comes from making wrong answers cancel
A qubit register holds every answer at once with an amplitude each; algorithms steer the amplitudes.
What does a quantum computer actually do that an ordinary one cannot? Here is the smallest real example: find one marked item among four. Two qubits hold all four possible answers at once, each with an amplitude of one half. Measure now and you get a random one. The first operation marks the right answer by flipping the sign of its amplitude. Measure now and it is still random: a sign is invisible to measurement. The answer is hidden in the phase. The second operation reflects every amplitude about their average. The three wrong ones were above the average, and land on zero. The marked one was far below, and lands on one. Measure now and you get the right answer, every time, after one step. An ordinary search needs up to four.
How quantum computing works · Going deeper · 1 of 2
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
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
Today’s machines are noisy; error correction spends about 1,000 physical qubits per reliable one.
So how close is the machine that breaks encryption? The scale here multiplies by ten at each line. Qubits are fragile and make errors about once in a thousand operations. To get one reliable qubit you combine about a thousand physical ones. The number of reliable qubits demonstrated so far: single digits to low tens. The largest machines have a few thousand physical qubits. Breaking the common RSA key size is estimated to need about twenty million. That is a gap of roughly ten thousand times. It may close in ten years, or thirty, or not at all in this form. Nobody knows. And that uncertainty is exactly why the encryption is being replaced now: you cannot go back and re-encrypt what was already recorded.
Quantum code-breaking distance · Going deeper
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
Specialised silicon: chips built for one dataflow, such as matrix multiplication, and nothing else.
Now the hardware under AI. Take one fixed job: classify a single image. How much energy does it take? Each line on this chart is ten times less. A server processor: about one joule. It can run anything, and pays for that. A graphics processor: a tenth of that. Then the specialists. A chip that only multiplies matrices: thirty millijoules. The neural chip in a phone: ten. A custom chip for language models: five. And laboratory estimates for analog designs reach a tenth of a millijoule. Ten thousand times less than where we began. Multiply any of these gaps by ten billion requests a day and it is the difference between one power station and none. That is why chips are fragmenting into specialists.
AI accelerators · Going deeper · 1 of 2
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
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
The von Neumann bottleneck: memory and compute sit apart, and the trip between them dominates.
Why is there so much energy to save? Because of where conventional chips spend it. Since the first lesson on machines, memory has been in one place and the processor in another, with a bus between them. Fetching one number from memory costs about six hundred and forty picojoules. Multiplying with it costs one. The bars are to scale: the second one is that sliver. Almost all the energy goes on moving the number, not using it. So the radical designs put memory and arithmetic in the same place, as the brain does. And they borrow a second trick from it: spikes. A neuron that has nothing to say stays silent and uses almost nothing. Your brain runs eighty-six billion neurons on about twenty watts.
Data movement and energy · Going deeper
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
Motion-to-photon latency: the time from a head movement to the light that shows it.
One more thing that is shipping: computers you wear on your face. They bring the tightest time budget in consumer computing. On a desk screen, thirty milliseconds of delay goes unnoticed, because your head is not attached to it. A headset is. When you turn your head, the picture must turn with it within about eleven milliseconds, or the mismatch makes people feel sick. That is too short to simply draw faster, so the system cheats, twice. First it predicts. It measures how your head is moving and aims the picture at where your head will be when the light comes out. Then it corrects. Just before display, it reads the head position again and shifts the finished picture to match. The application can even miss a frame, and you will not see it.
Spatial computing latency · Going deeper
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
PUE, power usage effectiveness: total power drawn divided by the power that reaches the computers.
All of this runs on electricity, and the amounts have changed character. First, how much of the power even reaches the chips? In an older data centre, about half. The rest is cooling and conversion losses. In a modern one, ninety-one percent. But the chips themselves got hotter. Air can cool a rack of about thirty kilowatts. A rack of AI accelerators draws eighty to a hundred and twenty. So liquid now runs across the chips. And the totals. Training one leading model in twenty twenty used about thirteen hundred megawatt hours. That is this sliver. By twenty twenty-five: fifty to two hundred thousand. The bar is to scale. Compute used to be a software cost, then a chip cost. It is now a question of power stations.
Computing and electricity · Going deeper
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
Hyperscale training: one job across 10⁵ chips, built around checkpoints and communication.
What is it like to run one job across that much hardware? Take ten thousand graphics processors. On an ordinary day, between half and two of them fail. The job runs for months. So failure is not an emergency: it is the weather. The whole state of the job is saved at regular intervals. When a chip dies, a spare is swapped in, and only the work since the last save is repeated. Minutes, not months. Now efficiency. Here is everything the hardware could compute if every chip ran flat out. A good training run uses thirty-five to fifty percent of it. The rest of the time, chips are waiting: for the network, for each other. Every step, all of them must exchange results. Past a point, adding chips mostly adds waiting.
Training at scale · Going deeper · 1 of 2
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
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
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?
However fast the 95% becomes, the 5% still takes one twentieth of the original time. 1 ÷ 0.05 = 20. That is Amdahl’s law.
Scaling laws
More compute has reliably meant more capability. So far
Scaling laws: model error falls as a smooth power law in compute, parameters and data.
Where is AI itself heading? Here the honest answer includes: nobody knows. Start with what is measured. For language models, error has fallen along a remarkably straight line as compute, model size and data grew ten-fold at a time. From GPT-2 in twenty nineteen, through GPT-4, to the models of today. Abilities nobody trained for, such as arithmetic and coding, appeared along the way. Does the line continue at the same slope? Or do returns shrink? Some labs report smaller gains per step, and the supply of good text is finite. Meanwhile a newer lever appeared: spend more compute on each answer, letting the model reason before it replies. Whether that is a new curve or a one-time lift, the next few years will show.
Scaling laws · Going deeper
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
The walls: light speed and the Landauer limit are absolute; heat, memory and cost bend at a price.
Finally, the limits. An honest view of the future starts with knowing which limits can move. Two are physics. Light takes at least twenty-eight milliseconds from New York to London, and no engineering changes that. It is why computing keeps moving closer to users. And erasing a bit has a minimum energy cost. Today’s chips use about a million times more, so that wall is real, and far away. Two are costs. Chips can shed about a watt per square millimetre before cooling gets exotic. And memory is so slow relative to processors that hundreds of instructions pass during one fetch. Both bend if you pay. The last is the algorithm’s own: work that cannot be split caps the speed-up. Only a different algorithm changes it.
Physical limits · Going deeper · 1 of 2
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
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
A company with unlimited money wants to push back one of these limits. Which one can its engineers actually move?
The memory gap bends with money: stacked memory, bigger on-chip caches, computing inside memory. Light speed and the Landauer limit are set by physics.
Summary
What each layer is made of changes. The map stays
You now have the whole map: from one bit to the edge of what can be built.
Here is the whole course in one picture. Bits at the bottom. Machines that process them. Networks, and the security that makes them trustworthy. Systems, data and AI. And interfaces, where it all meets a person. Look at what stayed constant through sixteen lessons. The layers, and the contracts between them. The trade-offs: speed against cost against safety. And the limits. Now look at what the frontier changes. Not the layers. What each one is made of. New mathematics under security. New chips under AI. Perhaps, one day, something other than silicon under everything. And a bit is still a bit. You have the map. Every other course here takes one region of it and goes all the way down.
Lesson complete
You finished “Frontier”
Answer the questions along the way to see your score here.