Future Thinking Lab

Future Thinking Lab

Download on the App Store

Future Thinking Lab episodes

  • Glass Substrates: Hype vs Timeline

    Intel announced glass substrates for chip packages in 2023, and as of November 2026 they still aren't in a commercial product. Helen explains why glass appeals for huge AI packages, where it cracks, and sorts the timelines from Intel, SKC, Samsung and analysts into demonstrated, announced and expected.

    You'll learn:

    • Why flatness and thermal expansion matter under large chip packages
    • What makes through-glass vias and glass handling hard
    • How to read a technology timeline: who said it, when, and how firm it is

    For education only.

    6 min
  • Light Instead of Copper

    AI data centres move enormous amounts of data, and copper links cost power over distance. Helen explains co-packaged optics, which converts signals to light right beside the switch chip, what the companies claim it saves, and why lasers and fibres make servicing the open question.

    You'll learn:

    • Why longer electrical paths cost signal quality and power
    • How waveguides, modulators and external lasers fit together
    • What Meta's large test showed, and what it didn't

    For education only.

    6 min
  • The NPU in Your Laptop

    Many new laptops come with a neural processing unit and a TOPS rating on the box. Helen explains what an NPU actually does, why TOPS is a peak figure rather than a measured speed, and why memory often sets the pace for running AI models locally.

    You'll learn:

    • How an NPU uses multiply-accumulate arrays and low-precision math
    • Why TOPS figures are hard to compare across chips
    • Why memory size and bandwidth limit local chatbots

    For education only.

    6 min
  • You Can't Fix What You Can't Measure

    Measurement plays a key role in up to half of all chipmaking steps. Helen explains how fabs measure feature sizes, layer alignment and defects, and how statistical process control spots drift before chips fail. Plus why 3D transistors make measuring harder.

    You'll learn:

    • The difference between control limits and specification limits
    • How Western Electric rules catch drift, and what they cost in false alarms
    • How run-to-run feedback corrects a process batch by batch

    For education only.

    6 min
  • Gate-All-Around Transistors

    As transistors shrank, their gates lost grip on the channel and current leaked even when they were switched off. Helen traces the path from flat transistors to FinFETs to gate-all-around nanosheets, and explains how nanosheets are built layer by layer. She also covers who ships them today and what imec sees coming next.

    You'll learn:

    • Why gate control, leakage and short-channel effects drive transistor design
    • The four new steps to build a nanosheet: superlattice, inner spacers, channel release and metal gate
    • How Samsung, Intel and TSMC versions compare, and what forksheet and CFET add

    For education only.

    6 min
  • Backside Power: Flipping the Chip to Feed It

    For decades, chips received power and signals through the same crowded wiring on top. Now chipmakers build the chip, flip the wafer over, and feed power from the back. Helen explains how it works, what Intel and TSMC claim it delivers, and the trade-offs in heat, debugging and cost.

    You'll learn:

    • Why front-side wiring got crowded, and what "IR drop" means
    • How a wafer is flipped, bonded, thinned and wired from behind
    • What Intel's PowerVia and TSMC's Super Power Rail claim, and what they cost

    For education only.

    6 min
  • High-NA EUV: Printing Smaller With a Wider Lens

    The newest chipmaking machines cost at least €350 million each, and their key upgrade is a wider lens. Helen explains how raising numerical aperture from 0.33 to 0.55 lets EUV light print finer features. She also covers why that halves the exposure field, and how Intel and TSMC are betting differently on when the trade-off pays.

    You'll learn:

    • The simple resolution rule behind every lithography machine, and which knob High-NA turns
    • Why anamorphic optics halve the exposure field, and what "stitching" means for big chips
    • The cost trade against multi-patterning, and Intel's versus TSMC's timelines

    For education only.

    6 min
  • Why AI Got About 280x Cheaper to Run

    In under two years, the price of getting GPT-3.5-level answers from AI fell more than 280-fold. In this episode of Future Thinking Lab, Helen uses Stanford's AI Index 2025 and Epoch AI data to explain where the savings came from. She covers smaller models, better chips, lower-precision maths and competition, and why total AI energy use keeps rising anyway. Sources are in the show notes.

    You'll learn:

    • What "inference" and "tokens" are, and how AI pricing is measured
    • The four forces behind the price drop, from small models to chip efficiency
    • Why cheaper AI per answer can still mean more electricity overall

    For education only.

    6 min
  • A Fab's Water Bill

    Making chips takes water purer than anything you've drunk, and a lot of it. In this episode of Future Thinking Lab, Helen explains why fabs rinse wafers in ultrapure water and how a leading chipmaker reports its water use and recycling. She also explains why water stress, like AI's electricity demand, is really a local story. Sources are in the show notes.

    You'll learn:

    • What ultrapure water is and why every chip depends on it
    • How fabs reclaim and reuse water, and how to read company recycling claims
    • Why location, not global totals, decides whether a fab strains local water

    For education only.

    6 min
  • How Much Electricity Does AI Actually Use?

    Data centres use about one and a half percent of the world's electricity. One AI data centre can draw as much as 100,000 homes. In this episode of Future Thinking Lab, Helen goes through the International Energy Agency's numbers. She covers where a data centre's power goes, why AI servers change the picture, and why the real strain is local, not global. Sources are in the show notes.

    You'll learn:

    • What uses the electricity inside a data centre, from servers to cooling
    • How much data centre power use the IEA expects by 2030 and 2035, and how wide the uncertainty is
    • How to read AI energy headlines: global vs. local, and which scenario

    For education only.

    6 min

About Future Thinking Lab

From the publisher's feed

Future Thinking Lab explores the ideas shaping how we think, work, and live in a rapidly changing world.