| Episodes drawn on | 327 |
| Cited here | 87 |
| Claims extracted | 108 |
| Gathering | concept union |
| Other themes | |
Organisational scale determines what an electronics company can build, what it can support, and which customers it can serve. The large instrument and semiconductor firms of the mid-twentieth century were internally subdivided so that engineering stayed close to its customers — Hewlett-Packard in the early 1980s ran as 82 divisions, each holding a product line and each operated as a self-contained small company.[144] Scale also permitted vertical integration and long development horizons that have no modern equivalent: Tektronix manufactured its own cathode ray tubes[421], and one Hewlett-Packard technology programme ran sixteen years before appearing in a shipping product.[232] Subsequent consolidation reorganised both industries into a small number of holding structures, which changed part availability, documentation, support routes and supply risk for every customer downstream.[152][404]
Divisional organisation
Under the divisional model most Hewlett-Packard divisions held 50 to 100 people, the largest — disc memory — about 300, and the arrangement kept engineering close to its customers even though the company looked fragmented from executive level.[144] Product development inside it ran on what was termed next-bench syndrome: an engineer built a tool to solve the problem of the engineer at the adjacent bench, which supplied an immediate local customer whose response validated or killed the idea before large sums were committed.[232] Long-horizon technology work was tolerated to a degree since lost, though individual projects could still be cancelled abruptly with two years of work behind them.[232]
The working environment was arranged to produce contact rather than left to chance. Bell Laboratories deliberately mixed electrical and mechanical staff along the same corridors at a time when most companies separated people by discipline, on the reasoning that unplanned conversation across specialities produced ideas neither group would reach alone.[165] A fixed daily ten o’clock coffee break and a Friday-afternoon beer bust were institutional features of Hewlett-Packard’s working culture, in contrast to individually partitioned modern workplaces.[165] Hewlett-Packard also gave engineers Friday afternoons for self-directed work decades before Google’s twenty-percent time made the arrangement famous.[57]
Selling engineer to engineer shaped the products themselves. Hewlett-Packard instruments carried bare numeric designations rather than marketed names because product naming was not routed through a marketing function.[144] The company ran and published formal research into the optimal legend colours for contrast and fast reading on the front panel of the HP-35 calculator, whereas later instrument front panels have been coloured on far thinner grounds, including a waveform button made blue because the designers preferred blue.[72] Internal standards, architectures and user interfaces were developed in-house, and when IBM’s personal computer appeared in 1982 Hewlett-Packard engineers judged it less capable than machines the company already had and declined to follow; building to the IBM specification came only after that architecture had become dominant, and the correction took years.[99]
Employment terms at that scale reached the engineer’s own work. Hewlett-Packard contracts claimed first rights to inventions made by employees, so Steve Jobs, then working in the calculator department on calculator chips, was obliged to offer the Apple I to Hewlett-Packard before building it himself; the company turned it down.[506] Hewlett-Packard itself began in a Palo Alto garage with a piece of test equipment, the 200A audio oscillator, as its first product, and is the origin of the garage-startup template later applied across Silicon Valley.[369] Hiring at the large semiconductor firms tracked business cycles rather than individual need: a 1974 graduate recruited by National Semiconductor was laid off after about two months, the company having spent more relocating him to California than it paid him in wages.[185]
The cost structure of a large organisation is easy to underestimate from outside. An internal analysis at one large instrument and medical-device firm put the fully burdened cost of shipping an empty cardboard box at about three hundred dollars.[302] Internal separation also propagates to the customer: Texas Instruments is organised into strongly separated vertical business units with sensor teams and processor teams working independently, so every chip carries its own reference design and the board designer becomes the integrator of the surrounding catalogue parts.[163]
Vertical integration
Tektronix was vertically integrated to the point of manufacturing its own cathode ray tubes, and supplied instruments with ceramic terminal strips and matching silver solder in the box, so a period instrument exposes the entire manufacturing stack of the company that built it.[421] The performance gap between incumbent oscilloscope makers and low-cost entrants still rests on silicon: Tektronix and Agilent develop custom acquisition ASICs, while low-cost USB instruments are assembled from commodity parts and advance only as fast as commodity silicon does.[209] Those in-house ASICs stay in service for decades — Tektronix has been reported to still use an ASIC designed thirty years earlier — because scope designers keep older silicon on mature processes rather than requalify a signal path.[646] Production testing of analog parameters on analog parts was likewise done with homemade testers rather than universal commercial machines: racks of Hewlett-Packard instruments under GPIB control, assembled and programmed in-house for each part.[348]
Integration was sometimes forced by the absence of anything to buy. An early-1980s PCB CAD company such as Cadnetics had to design its own workstation hardware as well as write the software, because no standard machine could run that class of code; the window in which such integration was necessary was short, and firms committed to it were stranded when standard machines arrived.[99] The opposite transition, from the bit-slice 2900 family to the single-chip CMOS 29000, replaced a per-design instruction set with a fixed one, which is what allowed standard compilers, in-circuit emulators and a supporting tool ecosystem to exist at all, the CMOS process also running far cooler than the bipolar parts.[103] Laser printers from Hewlett-Packard and Apple were the volume applications that briefly made the AMD 29000 one of the highest-shipping CPUs, processor volume of that era coming from office peripherals rather than computers.[103]
In semiconductors, owning fabrication remains a structural advantage. Texas Instruments still runs its own fabs, and a vertically integrated vendor aggregates its own product lines into internal scale at the wafer level, so any one order is a slice of a much larger run — something a fabless competitor bidding for foundry capacity cannot reproduce.[628] The effect shows up in allocation: in the 2008 shortage Linear Technology, which ran its own fabs, quoted lead times around thirty weeks on parts for which Texas Instruments was quoting ninety-nine.[628] Such capacity is financed on long horizons, Texas Instruments borrowing against an expectation of five to six billion dollars in annual revenue once its expansion is fully operational, a payback measured in years rather than product cycles.[578] Wafer-size transitions do not lower part prices for the buyer, because the capital cost of the new fab — on the order of seven billion dollars for the 450 mm generation — is recovered through the parts it produces.[63] New investment is also aimed at leading-edge nodes, while the parts that went unobtainable in the shortage were made on mature processes and older wafer sizes whose domestic capacity had moved offshore once it stopped paying for itself, so rebuilding leading-edge capacity does not restore their availability.[573]
Instruments, service and the installed base
Signature analysis was Hewlett-Packard’s attempt to systematise digital troubleshooting: known-good behaviour was captured by running dedicated test software on the instrument, and a logic probe at a defined node returned a short four-character signature the technician compared against the documented value.[169] Hewlett-Packard test equipment built from roughly the mid-1970s to the mid-1990s carries these codes internally, so the technique remains available across that whole generation of instruments to anyone who can obtain the data.[169] The codes are worthless without the service manual generated against that specific instrument, since a signature is meaningful only against the recorded known-good value, and losing the documentation strands a diagnostic capability built into the hardware.[169]
Custom and relabelled silicon has the same effect. Instrument makers specified semi-custom devices under house part numbers, as in the Tektronix TDS220 whose custom part was manufactured by National Semiconductor; no public data exists for such devices and a repairer can at best identify the standard part the design derives from.[227] Companies buying in Hewlett-Packard’s and IBM’s volumes had semiconductor makers silkscreen house numbers onto otherwise standard parts, so a board can carry a commodity device such as a 74HC00 under an in-house number, and a company running the scheme for decades can exhaust its own number space.[445] Vendor consolidation removes the service organisation as well: factory parts support for TM500-era Tektronix equipment has ended, so keeping that surviving instrument base alive depends on scavenging or fabricating parts.[655]
Instrument market structure is visible in pricing. The HP 8510 was the reference network analyser of its era at roughly a quarter of a million dollars fully loaded, and enough of a platform that third parties built and exported add-on boxes for it, including systems for measuring the radar signature of stealth aircraft.[214] Entry-level oscilloscope pricing held around seven to eight hundred dollars across thirty years of magazine advertising, a floor held deliberately because a vendor dropping below it undercuts the rest of its own range.[186] The incumbents effectively withdrew from the low-end oscilloscope market once cheap capable instruments became widely available, Tektronix having previously served that segment with rebadged product such as the TDS-220 rather than its own design.[347] Even the dominant oscilloscope manufacturer distributed other companies’ instruments through its national subsidiaries where its own range did not cover a segment, Tektronix Australia selling a competitor’s scope alongside its own.[130] The secondhand market for current-model instruments is thin and badly priced: a 300 MHz four-channel Tektronix DPO4034 with probes plus a Fluke 87, a package listing near ten thousand dollars new, changed hands privately for five hundred dollars against a resale value of six to seven thousand.[18]
Procurement structure inside large customers feeds back into how instruments are sold. Capital expenditure thresholds require sign-off above a fixed amount, so a buyer takes the cheapest instrument that clears the limit and adds bandwidth or features later as separate smaller purchases, which is why instrument vendors sell software-unlockable upgrades.[145] Architectural inheritance limits what those upgrades can deliver: spectrum functions grafted onto an oscilloscope architecture inherit its sweep behaviour, and on a Tektronix MDO3000 in RF mode narrowing the resolution bandwidth to 300 Hz or below over a low-megahertz span stretches a single update out to tens of minutes.[209]
Design decisions at these firms set grade boundaries that persist. Hewlett-Packard engineered its calculators to be correct in the final displayed digit while Sinclair deliberately accepted approximately right results to reach a low price point, and the same accuracy-versus-cost split separates instrument grades generally.[361] The Texas Instruments calculator division sustained margins far above the rest of the company, selling for around one hundred and fifty dollars a product costing roughly five to build, which left it with discretionary money other divisions did not have.[164] The HP-15C of the Voyager landscape-format series shipped from about 1982 until the end of the decade and was later reissued after a sustained campaign by users, so a discontinued design can return when its user base is organised enough to make the demand visible.[53] Hewlett-Packard has also released one of its older calculator designs as open source, publishing schematics and supporting design material.[36] Tektronix’s late-1970s one-gigahertz purely analog oscilloscope originated as a request from the United States Atomic Energy Commission for nuclear weapons testing, agreed on terms of thirty thousand dollars per unit — then the price of a house — a thousand-unit commitment from the customer, and the right to sell the resulting instrument on the open market afterwards.[119] The earliest Tektronix instruments, the 511 through 514, were grey, and the blue that became the company’s signature arrived with the 535 series; touch-up spray paint in that specific blue is still sold for restoration work.[119] Trade dress is defended in the same market: threatened by Fluke’s lawyers over the yellow multimeter holster, Tektronix changed a meter’s holster from yellow to blue rather than test the claim in court, leaving two otherwise identical meters distinguished only by holster colour, and the claim itself has never been adjudicated.[190]
Consolidation
The decision to break Hewlett-Packard in two was financial rather than technical: computer businesses commanded higher stock-market multiples than instrument businesses, so separating them was expected to raise the valuation of the computing half.[144] The split took effect in 1999, the computer and printer business retaining the Hewlett-Packard name while the original test-and-measurement business the company had been founded on was renamed Agilent and later Keysight.[164] Hewlett-Packard subsequently announced it would divest its personal-computer business and step back from consumer products, the reasoning being that consumer-level margins were not worth competing for.[57]
Test and measurement consolidated in parallel. Tektronix, Keithley and Fluke came under common ownership by Danaher, placing much of the mid-range market inside a single holding structure.[152] Assurances that an acquired instrument company will keep its own culture are testable against what happens to its site and brand, and the Keithley facility was rebranded Tektronix after the acquisition.[310] Tektronix’s Beaverton campus was progressively sold off as the company contracted, Xerox buying the printer operation and moving onto the same campus while Nike occupies other former buildings, still referred to by their Tektronix building numbers.[149]
The semiconductor lineage runs from Shockley through Fairchild: Noyce and Moore left Fairchild to found Intel, National Semiconductor came from the same lineage, and Linear Technology in turn spun out of National.[313] Those firms were funded and governed by a small overlapping group — one co-founder of National Semiconductor was also involved in funding Linear Technology and chaired the boards of Cypress, Microchip and Vitesse.[129] The physical legacy remains: the former National Semiconductor site in Silicon Valley is now a contaminated brownfield rather than an operating plant.[177]
Absorption leaves traces in the documentation and the catalogue. After Texas Instruments absorbed National Semiconductor the National data sheets were reissued under TI branding, typically by scripting a TI cover page onto the original document, so documents circulate in hybrid form with National URLs at the foot of pages carrying TI logos at the top.[102] Post-merger layoffs followed from portfolio overlap rather than financial distress, consolidation eliminating duplicated roles across two catalogues addressing the same markets.[184] Chip design headcount is the last thing cut in such a consolidation, because analog and silicon design skill disperses irreversibly once released and cannot be rebuilt on demand.[184] Texas Instruments issued last-time-buy notices covering roughly a hundred parts tied to one old National plant, including long-lived audio devices such as the LM386, parts of that age ending not because demand disappears but because the specific fab that makes them closes.[271] Consolidation in analog silicon has left effectively two large independent vendors after Analog Devices absorbed Linear Technology, so a jelly-bean function such as a charge-pump inverter that once had many independent second sources now has variants under only a couple of corporate roofs.[404] Merging vendors also correlates supply risk that used to be independent: where a surviving catalogue runs through common processes and fabs, a single problem propagates across every product line at once, whereas separate companies with different processes failed at different times.[628]
Acquisition changes the acquired company’s leadership as well. After Texas Instruments acquired ChipCon for a couple of hundred million dollars, the founder ran the resulting TI division for about a year before leaving, a common pattern where the person suited to building a small company is not suited to operating a division of a large one.[95] Corporate research output should be read with the same caution: Hewlett-Packard developed the memristor and pursued commercialisation with a Korean memory manufacturer, but a research announcement becomes a product only once the part can be ordered from a distributor.[7]
Vendor engagement across customer sizes
Access to a semiconductor vendor is a function of volume. At the largest vendors, design-in attention is reserved for very high-volume customers such as Dell and Hewlett-Packard, or the contract manufacturers building for them, while smaller designers interact with the catalogue instead.[122] Vendors differ structurally in whom they can serve: some deliberately build only for tier-one accounts, and Texas Instruments built application and sales structures to support customers of any size, the problem case being a vendor that retains a tier-one focus but sells to smaller customers anyway without the structure to support them.[452] Service scales with what a customer can spend, though a startup can still obtain samples and field-engineering attention by making its product into a story the vendor’s own marketing can retell later.[385]
Before the internet the vendor data book was the only design information available, so vendors that handed data books out freely or sold them through local shops were designed into products while those that restricted them lost sockets — the mechanism behind the long-lived preference for National and TI parts among engineers of that generation.[52] Access was otherwise mediated by field sales representatives, so obtaining a data sheet or a sample could require negotiating with an intermediary and was gated on projected volume.[52] Texas Instruments once ran a single large technical conference that functioned as a week of concentrated electronics education framed as a sales event, and it determined the part choices of attendees for more than a decade afterwards.[442] The format that made it effective was one engineer per part who had actually designed a product with it, present to answer questions and issue development kits; splitting or dropping such an event removes the principal route to introductions and documentation for parts that are otherwise hard to learn.[442] The current equivalent is staffed differently: Texas Instruments places the applications engineers who design and work with the products on its public E2E forums, some spending ten to twenty hours a week answering there, which makes the forum rather than the sales channel the practical support route for a small customer.[212] New silicon was once introduced to large companies by sending factory staff to run classes for their assembled engineering teams, and Hewlett-Packard, Tektronix, NCR and Ford were taught what a microprocessor was and how to wire one up in exactly that format.[241]
Documentation is filtered on projected support cost. A chapter covering the programmable real-time units was removed from a Texas Instruments technical reference manual on management review as likely to generate too much technical support, then republished separately through BeagleBoard with the explicit caveat that TI would not answer questions on it.[378] A silicon vendor can also withdraw software support while the part remains in production, forcing designs already committed to a processor off the vendor SDK.[325] Third-party board-support vendors fill that gap, covering most of a processor catalogue with better documentation and paid support at roughly five thousand dollars a licence, which for a small team amounts to buying an employee.[325]
Vendor tooling is pitched at one audience and cannot satisfy both: the simplified WebBench power-supply flow gave newcomers an answer in minutes while frustrating experienced designers who wanted inductor data, topology comparison and waveforms of the kind SPICE provides.[392] Training material written for internal use, such as a four-part Texas Instruments course on op-amp stability, circulates publicly and is substantive design instruction rather than promotional material.[341] Vendor-sponsored reference stacks accelerate a proof of concept only while the design stays on the exact listed module and processor; substituting a neighbouring part number or adding one unsupported requirement returns the team to greenfield integration, often after the schedule has been committed on the strength of the demonstration.[370] Where a part implements a published standard such as USB Power Delivery, development information issued by one vendor applies directly to a competitor’s controller, and code written against the standard rather than the vendor library survives an end-of-life part swap.[449]
Purchasing terms are similarly stratified. Negotiated part pricing still applies at hundred-thousand-piece volumes and is unavailable below that, and customers who must buy a full reel to obtain it commonly push the unused inventory onto their assembler to hold.[411] Semiconductor vendors will trade discounts on the order of fifteen to twenty percent for cash paid up front, since a supplier will always take cash today over cash later, so a cash-rich customer uses prepayment as an investment vehicle as much as a purchasing arrangement.[240] During allocation, offering higher volume does not move a small customer up the priority list: lead times are quoted flat — two months in one case — with the instruction to plan further ahead, because the vendor is scheduling wafer starts rather than shipping from distributor stock.[389]
Internal capability displaces purchased services above a certain size. Large companies build their own component libraries and footprint data rather than buying it, which caps the addressable market for selling such data to precisely the customers least able to pay.[118] Large chip companies design on the commercial EDA tools, principally Cadence and Mentor, so experience with open-source ASIC design software counts as familiarity with the flow rather than as a matching skill when applying.[141] Above roughly fifty thousand machines, buyers stop purchasing conventional 19-inch enterprise servers from Hewlett-Packard or Dell because the price carries plastics, front panels and per-machine status displays that serve no purpose in a room nobody visits, and at that scale the specification is rewritten around what the operator actually observes.[357]
Parts, standards and long-term supply
Recent integrated parts arrive in vendor-proprietary packages that competitors are not permitted to copy, so there is no drop-in second source and a single part choice carries far more schedule and supply risk than an equivalent decision in the era of standard packages.[244] Exclusivity contracts that bar a vendor from selling a given part to anyone but one customer are a deeper obstacle to independent repair than withheld schematics or service tools, because no amount of documentation makes an unpurchasable part replaceable.[594] Package markings are not evidence of provenance in either direction, since counterfeit parts are silkscreened with genuine vendor branding, the marking cost being negligible against the price of the part being imitated.[351] Where integrated parts become unobtainable and automotive-qualified stock disappears first, one response for a design that is not space constrained is to trade board area back for availability, rebuilding a function such as a switching converter from a small microcontroller, an op-amp and a pass transistor that can still be sourced.[558]
Long-life programmes impose their own contractual structure. Military and long-life industrial work requires written commitments from semiconductor suppliers that a part will still be supported in ten to fifteen years, and extracting those statements is slow and difficult even when ultimately possible.[366] Long-term parameter drift is specified only on selected premium products, and an ordinary op-amp carries no long-term stability specification at all, which matters when the equipment it sits in must hold calibration for a decade in an industrial installation.[366] Substituting a different manufacturer for an electrically identical part must still be recorded against the build, because lot-level differences only become visible as field failures and are untraceable without that record; high-end CAD tools support approved alternates so purchasing can choose freely within a vetted set.[178]
Intellectual property shapes the catalogue as much as process technology does. Expired patents on the 8051 architecture allow any vendor to embed the core without a licence fee, which is why it persists inside modern radio system-on-chips such as the Texas Instruments CC2540 long after it stopped being competitive on its merits.[326] Two opposing positions coexist inside large semiconductor firms: patenting publishes the method and lets competitors design around it, so a technique may be better hidden in the silicon, while top-down pressure to generate patentable work persists because patent counts are presented to shareholders.[270] Standards bodies are contested by vendors trying to get proprietary features written into the specification so that their silicon becomes the compliant choice, with tactics extending to bringing extra members to voting sessions and holding pre-meetings to set strategy.[245] Trademark reaches product vocabulary too: Texas Instruments held the trademark on the term sprite, so Commodore’s engineers were required to call the same graphics feature a movable object block.[222] Printer makers authenticate consumables with a chip in the cartridge, and third-party cartridges sold as chipped are still detected as non-genuine and produce a warranty warning though they continue to function.[303]
Segment reporting and pricing history record where the large vendors have concentrated. Major semiconductor manufacturers report digital power control, and often motor control with it, as a distinct revenue segment, a direct signal of how large those markets have become relative to the general catalogue.[9] Texas Instruments released the 5400 logic family in 1964 for military use, the commercial 7400 equivalent following about two years later, with integrated circuits not appearing in hobby and commercial magazine projects until 1967.[207] Motorola priced the 6800 microprocessor at about two hundred and fifty dollars with its peripheral interface adapter at about five, and engineers at the companies visited wanted the technology while their employers concluded that at that processor price they could not build a product around it.[241] Texas Instruments made high-power 900-nanometre gallium arsenide infrared LEDs at Richardson, north of Dallas, in the era when they were new, and an interested outsider could reach the responsible research scientist directly by writing to the company.[171]
Small organisations
Small companies carry risk that a division of a large one absorbs. Before microprocessors were available, a small company could commission a full-custom integrated circuit from National Semiconductor for about one hundred thousand dollars — in one case a 200-gate-equivalent part — and the programme was delivered a year late, which for a small company was fatal to the product it was intended for.[109]
Process discipline is where the two structures differ most visibly. Informal environments substitute drive-by opinion for structured design review, and a passer-by proposing a different switching regulator weeks before an engineering-validation build carries none of the schedule risk of the change, which is the specific defect a formal review with named owners removes.[394] An engineer trained through Hewlett-Packard and then on FAA- and FDA-certified products treats prototype work assembled from modules as categorically different from production work, holding that a qualified electrical engineer must vet the system before volume rather than letting a working prototype pass as a design.[281] Some capability has moved the other way: large electronics companies increasingly spin and hand-build prototype boards internally rather than through a contract line, with the parts involved pushing stencil requirements toward cut features in the tens of microns.[320]
Growth itself is treated as a decision rather than a default. A growing hardware company faces an explicit choice between adding business units and keeping the working conditions that produced its output; SparkFun addressed the question deliberately as a company, with Hewlett-Packard’s trajectory as the reference case, and capped growth rather than lose the environment.[157]
Large employers also seed the small companies around them. Regional engineering clusters persist through employees leaving an anchor employer, taking the proceeds and starting a firm nearby because they already live there, a pattern seeded by the aerospace and instrument anchors around Everett and Beaverton.[421] Recruitment follows the same concentration: Fluke advertised for engineers in the Denver and Boulder area near Hewlett-Packard’s Loveland division and hired on the spot, the specialised talent pool already being there.[180]