Locara

A Brief History of Chips and Computers — From Bell Labs to LLMs

What this is: The arc from the first transistor (1947) to today’s local-AI hardware (2026), focused on the inflection points and the people behind them. Compressed but with enough specificity to ground the rest of the chip / LLM notes. Why it matters: Most of the dynamics shaping local AI today (Moore’s Law’s slow death, vertical-integration cycles, the open-vs-closed pendulum, distribution wins) have direct precedents earlier in computing history. Knowing the cycles makes Locara’s strategic decisions less ahistorical. Most relevant to Locara: Context. Pairs with chip-fundamentals.md for the technical model and modern-chip-landscape.md for the current product matrix.

1. Pre-transistor: theory and early machines (1936–1947)

  • Alan Turing (1936) — On Computable Numbers. The Turing machine and undecidability; the theoretical foundation. Cracked Enigma at Bletchley Park during WWII.
  • Claude Shannon (1937 master’s thesis) — Boolean logic ↔ electrical relay circuits. The link from logic to switching that everything else stands on. Later (1948) founded information theory.
  • John von Neumann (1945, First Draft of a Report on the EDVAC) — stored-program architecture: code and data in the same memory. The “von Neumann architecture” that virtually all general-purpose computers still use.
  • ENIAC (1945, J. Presper Eckert and John Mauchly, U. Penn) — vacuum-tube computer, ~17,000 tubes, programmed by rewiring.
  • Colossus (1943–44, Bletchley Park, Tommy Flowers) — earlier than ENIAC for code-breaking; classified for decades.

These machines used vacuum tubes — big, hot, fragile. The transistor was the unlock.

2. Transistor era (1947–1958)

  • December 1947, Bell Labs: John Bardeen, Walter Brattain, William Shockley demonstrate the point-contact transistor. Shockley invents the bipolar junction transistor in 1948. Nobel Prize 1956. Jon Gertner’s The Idea Factory is the canonical history of Bell Labs as an institution.
  • Shockley Semiconductor Lab (1956) — Shockley moves to Palo Alto, hires brilliant engineers. His management is so toxic that…
  • The Traitorous Eight (1957) — eight of his engineers (Robert Noyce, Gordon Moore, Eugene Kleiner, Jean Hoerni, Jay Last, Sheldon Roberts, Victor Grinich, Julius Blank) leave to found Fairchild Semiconductor with Sherman Fairchild’s backing. This single act seeds Silicon Valley.
  • Jean Hoerni invents the planar process at Fairchild (1959) — flat photolithographic transistor manufacturing, the foundation of all subsequent IC fabrication.

3. Integrated circuit (1958–1971)

  • Jack Kilby (Texas Instruments, 1958) demonstrates the first IC — multiple components on one piece of germanium. Robert Noyce (Fairchild, 1959) independently develops a more manufacturable silicon planar IC. Both credited as IC inventors. Kilby gets the Nobel Prize (2000); Noyce had passed in 1990.
  • 1965: Gordon Moore publishes Cramming More Components onto Integrated Circuits in Electronics magazine — “Moore’s Law”: the number of transistors on a cost-effective IC doubles approximately every two years. Originally observation of a 4-year trend; held for 50+.
  • 1968: Noyce + Moore leave Fairchild to found Intel. Andy Grove joins shortly. The original Intel.
  • 1971: Federico Faggin (Intel, leveraging the Italian silicon-gate technology he’d developed at Fairchild) leads design of the Intel 4004 — the first commercial microprocessor, a 4-bit CPU on a single chip, ~2,300 transistors. Ted Hoff and Stanley Mazor co-credited on architecture.
  • Robert Dennard (IBM, 1968) invents DRAM. Dennard scaling (1974 paper) — the empirical observation that as transistors shrink, they get faster and more power-efficient simultaneously, so power density stays constant. Held until ~2005.

4. Personal computing era (1975–1995)

  • 1975: MITS Altair 8800, Intel 8080. Bill Gates and Paul Allen write a BASIC interpreter; Microsoft (then Micro-Soft) is founded.
  • 1976–77: Apple I, II. Steve Wozniak’s design, Steve Jobs’s company. Apple II (1977) becomes a cultural milestone.
  • 1981: IBM PC. Built on Intel 8088 + Microsoft DOS. The architecture was open enough — IBM didn’t own the BIOS exclusively → clones (Compaq, 1982) → x86 PC ecosystem.
  • 1984: Apple Macintosh. GUI from Xerox PARC’s lineage (Alto → Lisa → Mac path). Michael Hiltzik’s Dealers of Lightning tells the PARC story; Steven Levy’s Insanely Great tells the Mac one.
  • 1985: ARM at Acorn Computers (UK). Sophie Wilson designs the instruction set, Steve Furber the chip. Originally for the BBC Micro successor. Reduced Instruction Set, low transistor count, low power. Spun out as ARM Holdings 1990; the architecture eventually defines mobile computing.
  • 1985: David Patterson & John Hennessy publish papers establishing RISC as a coherent design philosophy. They later share the Turing Award (2017).
  • 1987: Morris Chang founds TSMC (Taiwan Semiconductor Manufacturing Company). The “pure-play foundry” model — Chang separates design from manufacturing. This is the birth of the fabless industry. Chris Miller’s Chip War is the canonical history.
  • 1991: Linus Torvalds posts Linux 0.01. A Unix-like kernel, free software. Becomes the substrate of nearly all server, mobile, and embedded computing.
  • 1993: Pentium ships. 1994: PowerPC ships (Apple/IBM/Motorola alliance), in early Macs.
  • The dot-com boom (1995–2000) — driven by cheap commodity x86 servers + the web (Tim Berners-Lee’s WWW, 1989–91, public 1993).

5. Mobile and the rise of ARM (2000–2010)

  • 2001: iPod — Apple as a hardware company first, before the phone.
  • 2007: iPhone — iPhone 1 ran an ARM11 core. The iPhone reshapes consumer computing toward mobile, ARM, and integrated SoCs.
  • 2008: Apple acquires P.A. Semi (chip design firm) — sets up Apple’s in-house chip ambition. Hires Jim Keller (then with AMD K8 fame).
  • 2010: Apple A4 in iPad and iPhone 4 — first Apple-designed SoC. Begins the trajectory toward Apple Silicon.
  • 2007: Android open-sourced (Andy Rubin’s project, acquired by Google 2005).
  • Qualcomm Snapdragon becomes the Android flagship workhorse through the 2010s.

Carver Mead and Lynn Conway’s Introduction to VLSI Systems (1980) was the textbook that made design-fab separation (and thus the fabless industry) viable. Mead also coined the term “Moore’s Law.”

6. The GPU and the deep-learning awakening (2006–2017)

  • 2006: NVIDIA releases CUDA. General-purpose GPU programming. Jensen Huang’s foresight predates the deep-learning era by ~6 years; the bet pays off in 2012.
  • 2009–2012: Deep learning starts working at scale. Geoffrey Hinton’s group, Yoshua Bengio’s, Yann LeCun’s (the trio shares the 2018 Turing Award). Restricted Boltzmann Machines, dropout, GPUs as training accelerators.
  • 2012: AlexNet (Krizhevsky, Sutskever, Hinton) wins ImageNet by a wide margin. Trained on two GTX 580 GPUs. The moment GPUs became the AI hardware.
  • 2014–15: GANs (Goodfellow), VGG, ResNet (He et al.). Deep learning eats vision.
  • 2017: Transformer. Vaswani et al. (Google), Attention Is All You Need. The architecture underlying every subsequent LLM. Originally for translation.
  • 2015–2018: TPU at Google. Norm Jouppi et al. — Jonathan Ross was a key engineer (later founds Groq). Internal-only at first, then GCP-available.
  • Dennard scaling broke around 2005; multicore became necessary as single-thread perf gains slowed.
  • 2010s vertical-integration era: Apple Silicon comes together (M1 in 2020), Google designs TPUs and later Tensor SoCs, Amazon designs Graviton (2018+). Hyperscalers go fabless-design.

7. The LLM era (2018–2026)

  • 2018: GPT-1 (OpenAI). 117M params. Modest.
  • 2018: BERT (Google). Bidirectional encoder; dominates NLP benchmarks.
  • 2019: GPT-2 (1.5B). OpenAI delays full release citing misuse risk; the move is later seen as the start of the closed-frontier philosophy.
  • 2020: GPT-3 (175B). The first model that felt qualitatively different. Language Models are Few-Shot Learners — in-context learning emerges from scale.
  • 2020: Apple M1. Apple Silicon ships in Macs. Unified memory + ARM at laptop performance levels. The local-AI hardware story begins to be plausible.
  • 2020–22: Scaling laws formalized. Kaplan et al. 2020 (“Scaling Laws for Neural Language Models”), then Hoffmann et al. 2022 (“Training Compute-Optimal LLMs” — Chinchilla). These papers reframe the field as a compute-and-data optimization problem.
  • November 2022: ChatGPT. The cultural moment. 100M users in two months. AI becomes consumer-mainstream.
  • 2023: GPT-4, Claude, Gemini. The closed frontier consolidates around three labs.
  • February 2023: Llama 1 leaks, then July 2023: Llama 2 ships with a permissive-ish license. The open-weights era begins in earnest.
  • 2023: NVIDIA H100 ramps. Export controls on H100 to China spark a major geopolitical realignment.
  • 2024: Multimodal everywhere. GPT-4o (May), Claude 3.5 (June), Gemini 1.5 (Feb). Apple ships Apple Intelligence (June 2024 announcement, late-2024 rollout) — local 3B + Private Cloud Compute hybrid.
  • December 2024: DeepSeek V3 ships, claimed $5.6M training cost.
  • January 2025: DeepSeek R1. Open-weights reasoning at o1 parity. Stock-market shock; widely-discussed proof that the closed-frontier moat is narrower than it appeared.
  • 2025: Reasoning models become standard. Open + closed both ship reasoning variants.
  • 2025–2026: Local AI viability inflection. A maxed Mac Studio runs 70–405B-class open models at usable speeds. iPhones run 3–7B Q4. The hardware-software combination Locara assumes finally exists.

Recurring patterns

A few cycles repeat across this history that are worth Locara’s attention:

  1. Vertical disintegration → reintegration. IBM owned everything (1960s); the PC era unbundled hardware/OS/apps; mobile re-bundled (Apple, Samsung); cloud re-bundled at the data-center scale (AWS); now AI is bundling again (Apple Silicon + Apple Intelligence; NVIDIA’s full stack from chips to CUDA to NIMs). Locara is making a bet on the next unbundling — that local AI breaks the cloud-AI bundle.
  2. Open ↔ closed pendulum. Every era has both. Closed wins early (better integration), open wins later (better economics, ecosystem). Mainframes closed; UNIX (eventually) open. iOS closed; Android relatively open. Frontier LLMs closed; Llama / DeepSeek open. Pattern: open catches up, then sets the floor.
  3. Software eats hardware moats slowly but surely. RISC-V is doing to ARM what Linux did to commercial UNIX. Open inference runtimes (llama.cpp, vLLM, MLX) are commoditizing what NVIDIA’s CUDA moat had previously protected.
  4. Distribution wins. Microsoft on x86 PCs, Google on Android, Apple’s App Store, npm, PyPI, Hugging Face. The platform that owns how things ship to users wins regardless of underlying tech. Locara’s bet is on this being the missing layer for local AI.
  5. The textbook is forever. Hennessy & Patterson on architecture, Goodfellow et al. on deep learning, Russell & Norvig on AI broadly — the canonical references hold for decades while the products churn. Locara should engage with the textbook level, not chase weekly news.

Specific learnings for Locara

  1. Cycles teach humility. Mainframe → mini → PC → mobile → cloud → local-AI is not a story of inevitable progress; it’s the wave we’re betting on. Be ready for cloud to recapture territory (e.g., if frontier capability gap stays large, the local-only positioning weakens). The best defense is making the local-only experience genuinely better, not just cheaper.
  2. Open eats closed in the long run, but not on day one. The frontier-closed phase of every era is real and lasts years. Locara should expect Claude/GPT/Gemini frontier to remain ahead of open weights for some years. Position accordingly: don’t try to compete on capability ceiling, compete on what closed cannot offer — privacy, ownership, durability, no-internet operation.
  3. The fabless / foundry split is the precedent for “framework / runtime / app” split. TSMC enabled a thousand fabless chip companies because it had the manufacturing economy of scale. Locara’s runtime + app store can play the analogous role for local AI apps.
  4. Standards matter more than products. x86, ARM, RISC-V, USB, HTTP, OpenAI’s API spec — the lasting wins are interface standards. Locara’s manifest format and app-runtime contract are the standards play; the registry and reference apps are the products. Invest in the spec.
  5. Engagement with primary sources. The Llama 3 paper, the DeepSeek V3 paper, FlashAttention papers, scaling-law papers, Hennessy & Patterson, Chip War. These should be on the Locara reading list and referenced in design docs. Don’t operate from secondary commentary.
  6. The “infrastructure team → company → enshittification” cycle has a long history (npm Inc., MongoDB Atlas pricing, Docker, Ollama Cloud). Locara’s structural decisions (OSS license, no-VC where possible, founder voice) need to be designed against it from the start.
  7. Personal computing is ~50 years old; it’s not done. Mobile is ~18 years old; ARM laptops are ~5 years old as a credible category; local-AI viability is ~2 years old. We’re in the early part of the local-AI curve. Plan for a long arc — Locara is not a 12-month project.

References

  • Walter Isaacson, The Innovators (2014) — broadest single-volume history, Babbage to Google. Best entry point.
  • Jon Gertner, The Idea Factory (2012) — Bell Labs.
  • Tracy Kidder, The Soul of a New Machine (1981) — Pulitzer-winning account of building a minicomputer. Conveys the engineering culture of the era better than anything else.
  • Michael Hiltzik, Dealers of Lightning (1999) — Xerox PARC.
  • T.R. Reid, The Chip (2001 ed.) — the IC’s invention and the Kilby/Noyce story.
  • Chris Miller, Chip War (2022) — geopolitics and economics; essential for understanding the post-2018 era.
  • Steven Levy, Hackers (1984), In the Plex (Google), Insanely Great (Mac). Levy’s broader oeuvre is a who’s-who of computing-history reporting.
  • Walter Isaacson, Steve Jobs (2011).
  • Brian Kernighan, UNIX: A History and a Memoir (2019) — the era of Bell Labs computer science from the inside.
  • Gordon Bell, Computer Engineering: A DEC View of Hardware Systems Design — the minicomputer era from a designer’s perspective.
  • Hennessy & Patterson, Computer Architecture: A Quantitative Approach — see chip-fundamentals.md.
  • Russell & Norvig, Artificial Intelligence: A Modern Approach — the AI textbook.
  • Vaswani et al., Attention Is All You Need (2017).
  • Kaplan et al., Scaling Laws for Neural Language Models (2020); Hoffmann et al., Training Compute-Optimal LLMs / Chinchilla (2022).
  • Computer History Museum (https://computerhistory.org) — interviews with most of the people named here. The oral history archive is a treasure.
  • Asianometry (YouTube) — accessible deep-dives on semiconductor history and current geopolitics.