TECH HISTORY — SATURDAY, SEPTEMBER 12, 2026

The Baloney Slicer: How a Plum Orchard in San Jose Gave the World the Hard Disk

Seventy years ago this month, IBM shipped 3.75 megabytes that weighed a ton. The machine it became still stores almost everything you own.

GadgetGlow Bytes Editorial

Published September 12, 2026 · 8 min read · How we work

In 1952 IBM sent a small group of engineers to a rented building on Notre Dame Avenue in San Jose, California, overlooking a plum orchard. The valley did not yet have its nickname. The team's brief was unglamorous and enormous: find a way for a computer to reach any record it needed without first chewing through every record in front of it.

Four years later they shipped the answer, and this month marks seventy years since the world found out about it. The IBM 350 Disk Storage Unit held 3.75 megabytes, stood five feet high by six feet wide, weighed more than a ton with its air compressor, and needed most of a room. Every hard disk drive manufactured since — including the 44-terabyte units going into hyperscale data centres right now — is a direct descendant of that cabinet.

The anniversary is worth more than a nostalgia photograph, because the interesting part of the story is not how small 3.75 megabytes looks today. It is that the disk drive has been declared obsolete roughly once a decade for seventy years and is currently enjoying its best revenue environment in history.

The Problem Nobody Could Solve: Getting There Fast

Before 1956, business computing ran on punched cards and magnetic tape. Both are sequential: to read record 40,000 you read the 39,999 records in front of it. For payroll that is tolerable. For inventory control, where a clerk needs one part number out of tens of thousands while a customer waits, it is useless. Retrieving a single piece of information could take hours or, in the worst filing arrangements, days.

The concept of random access was not new. What nobody had was an implementation fast enough to be worth building. IBM's San Jose lab, under Reynold B. Johnson — a former high school science teacher who would end his career as an IBM Fellow — worked through the options by elimination. According to IBM's own account of the project, the team tried rods, strips, tapes and flat plates before arriving at aluminium discs painted with iron oxide, read by an arm that hovered over the surface like a record-player needle that was never allowed to touch the record.

The team around Johnson included Louis D. Stevens, William A. Goddard and John J. Lynott, who filed the underlying “Data Storage Machine” patent on Christmas Eve 1954, and John Haanstra, who led the companion data-processing system. The Computer History Museum's record notes that the work was informed by Jacob Rabinow's notched-disk memory ideas published in Electrical Engineering in 1952 — the RAMAC was a synthesis, not a bolt from the blue.

Tech HistoryStorage

Four Problems, Four Improvised Fixes

What makes the RAMAC story unusually instructive is that each obstacle was solved by a fix that sounds like a workshop bodge and turned out to be permanent industry practice.

The discs warped. Spun at speed, a single aluminium plate deformed enough to ruin the geometry. The team glued two discs back to back, and the stiffened sandwich held its shape.

The head could not touch the surface. Contact would destroy data and the head alike. The answer was compressed air: the head was floated on a forced-air cushion. On the shipping product it rode roughly 800 microinches above the surface — about 20 micrometres, or a quarter of the thickness of a human hair — riding what the engineers described as a hydrostatic bearing. Seventy years later every drive on the planet still works by flying a head above a spinning surface; only the gap has changed.

One disc held nothing useful. So the team stacked them. An early configuration used sixteen discs mounted horizontally, and IBM executives who saw it reportedly nicknamed the contraption the baloney slicer. The production unit took the joke further: fifty 24-inch discs on a single spindle turning at 1,200 rpm.

The arm had to get anywhere, immediately. This was the hard one. Johnson later described the target plainly: travel from a track six inches into one disc, out, two feet down the stack, and six inches in again — in half a second. “We achieved something like 800 milliseconds,” he recalled, “and that's where the product came out.” The shipping specification was an average access time under one second. Against hours of tape handling, 800 milliseconds was not a compromise. It was a different category of machine.

Original graphic by GadgetGlow Bytes, compiled from the sources listed below.

June 1956: A Paper Company in San Francisco Goes First

The first production IBM 350 did not go to a laboratory or a government agency. Per the Computer History Museum's record, it was shipped in June 1956 to Zellerbach Paper in San Francisco, as part of the IBM 305 RAMAC system. The public announcement followed that September; contemporaneous trade accounts place it on 13 September, while IBM's own heritage page gives only the month.

The acronym has a small piece of revisionism attached. Researchers coined RAMAC from “random-access memory” plus the -AC suffix borrowed from the then-fashionable UNIVAC. Marketing later retrofitted it to Random Access Method of Accounting and Control, because accounting was where the demand was.

Orders followed quickly. IBM lists 3M, New York University, Norfolk Naval Shipyard, Pfizer and United Airlines among early customers. Within two years the US Custom House was using RAMAC to track Atlantic shipping, cutting response time on distress calls to five minutes. The machine went on tour, appeared in the US pavilion at the 1958 Brussels World's Fair, and handled scoring at the 1960 Winter Olympics in Squaw Valley, where calculations that had taken officials hours by hand came back in minutes. Roughly a thousand systems were built before the line ended in 1961.

On the numbers themselves, the record is not perfectly tidy, and it is worth saying so rather than picking the tidier figure. IBM's heritage page describes 5 million characters at 7 bits per character and a system storing “5 to 10 megabytes”. The Computer History Museum's specification sheet says 5 million 6-bit characters, equivalent to 3.75 megabytes, at an areal density of 2,000 bits per square inch. Both are credible institutional records of the same machine. We use the museum's figure where precision matters because it is stated with its encoding and its density, and we flag the disagreement because the internet's confident “5 MB” shorthand papers over it.

What Seventy Years of Density Actually Looks Like

The usual way to dramatise this is to hold up a microSD card. A better way is to draw the two platters at the same scale, because the shrinkage of the disc is the part people never picture.

Original graphic by GadgetGlow Bytes. Platter diameters drawn to relative scale; per-disc capacities derived from the sourced figures below.

Divide the museum's 3.75 megabytes across fifty discs and each 24-inch platter carried roughly 75 kilobytes. Seagate's current Mozaic 4+ platform puts 4.4 terabytes on a single 3.5-inch-class platter. That is an increase of about 59 million times per disc — achieved on a disc with roughly one-fortieth of the surface area.

Expressed as areal density the jump is larger still. The 1956 figure is documented: 2,000 bits per square inch. Neither Seagate nor Western Digital publishes a per-model areal density for current drives, so the modern figure has to be estimated. Taking a usable recording band between roughly 12 mm and 47 mm of radius on a 95 mm platter gives about 20 square inches of two-sided recording surface, which puts 4.4 TB per disc in the region of 1.7 terabits per square inch — on the order of 800 million times the 1956 density. That is our calculation from platter geometry, not a manufacturer specification, and should be read as an order-of-magnitude estimate.

Specification	IBM 350 (1956)	Standard nearline (2026)	Seagate Mozaic 4+ (2026)
Capacity	3.75 MB	30 TB	Up to 44 TB
Discs	50	10 (typical)	10
Disc diameter	24 in	~95 mm	~95 mm
Per disc	~75 KB	3 TB	4.4 TB
Recording method	Inductive, forced-air head	PMR / ePMR	HAMR (laser-assisted)
Spindle speed	1,200 rpm	7,200 rpm	7,200 rpm
Head flying height	800 microinches (~20 µm)	Not published per model	Not published per model
Areal density	2,000 bits/in²	Not published	~1.7 Tb/in² (our estimate)
Access time	Under 1 second (avg)	Single-digit milliseconds	Single-digit milliseconds
Mass	Over 1 ton (with compressor)	Under 1 kg	Under 1 kg
Cost basis	$750 per month, leased	~$668 purchase (Q1 2026 reporting)	Hyperscale contract pricing

Specifications compiled by GadgetGlow Bytes from manufacturer documentation and institutional archives, verified September 12, 2026. Nearline and Mozaic rows reflect published platform figures, not a single model datasheet; “typical” disc counts vary by SKU.

The cost comparison is where storage history gets sloppy, and it is worth doing the arithmetic rather than repeating the viral version. The museum records the 350 as leasing for $750 per month. That is roughly $200 per megabyte per month. To assemble a single terabyte you would have needed about 266,700 units — some $200 million per month in lease payments, or on the order of $12 billion over a five-year term, before you tried to find a building for 266,700 one-ton cabinets.

IBM's own heritage page cites a Computerworld figure of one trillion dollars per terabyte in the 1950s, against under $50 today. Both numbers make the same point, but they are about eighty times apart, and neither page explains the basis. If you have ever seen that trillion-dollar figure quoted as settled fact, this is a good reminder that storage-history statistics circulate faster than their footnotes.

59 million×

Capacity gain per disc

90%+

Cloud exabytes on disk

~22×

Flash cost premium per TB

100 TB

Stated 2029–2030 target

Why the Disk Refused to Die

Johnson's machine should have been a historical curiosity decades ago. Semiconductor memory was supposed to finish it. Flash was definitely supposed to finish it. Instead, according to Western Digital's own analysis of cloud infrastructure, more than 90% of the exabytes sitting in cloud data centres today are on spinning disks, with solid state holding the remaining tenth.

The reason is arithmetic, not sentiment. Industry pricing reported for the first half of 2026 put a 30 TB enterprise SSD at roughly 22 times the cost per terabyte of a 30 TB hard drive. When a hyperscaler needs to retain an extra exabyte of training data that will be read rarely but must never be thrown away, that multiple decides the purchase.

Which is why the recording head is now getting a laser. Heat-assisted magnetic recording briefly warms a spot on the platter during the write, allowing far more stable — and therefore far smaller — magnetic grains. In March 2026 Seagate announced that its HAMR-based Mozaic 4+ platform, supporting capacities up to 44 TB, was qualified and in volume production with two hyperscale cloud providers, on a stated roadmap from today's 4+ TB per disc toward 10 TB per disc and drives of up to 100 TB. Seagate's own figures put the efficiency case bluntly: across a one-exabyte deployment, it calculates roughly 47% better infrastructure efficiency than standard 30 TB drives, about 100 square feet less floor space and around 0.8 million kilowatt-hours less electricity a year.

Western Digital is taking a deliberately different route. At its February 2026 Innovation Day the company said its 40 TB UltraSMR drive — built on extended conventional recording rather than HAMR — was in qualification with two hyperscale customers for volume production in the second half of 2026, with its own HAMR drives ramping in 2027 and a stated path to 100 TB by around 2029. Its engineers have publicly sketched 14-platter designs reaching toward 140 TB in the 2030s. Two of the three remaining drive makers, in other words, disagree about whether lasers are needed yet.

Seventy years on, the fight is recognisably the same one Johnson's team had in the plum orchard: how much can you read back reliably from a surface spinning past a head that must never touch it.

Our Read

The anniversary coverage will frame this as a density story. The more useful reading is that the disk survived seventy years by repeatedly refusing to compete on speed. Drums, magnetic core, DRAM, then flash — each arrival took the performance tier, and each time the disk conceded it and retreated to the capacity tier, where it won on cost per terabyte. HAMR is that same trade made a fifth time. The laser is not there to make disks fast; it is there to keep them cheap at the exact moment AI workloads have made “retain everything forever” a business requirement. The vendor material frames HAMR as an AI story. We would frame it as confirmation that the roughly 22-fold cost gap between enterprise flash and nearline disk is the product, and the only number in this market that really has to hold. What the sources collectively skip is the arithmetic on the other side of the ledger: a 44 TB drive takes far longer to rebuild into an array than a 4 TB one did, because sequential transfer rates have not scaled anywhere near as fast as capacity. Density gains are being banked as cost savings while the rebuild window quietly widens.

What we still don't know: whether 10 TB per disc arrives on the stated schedule — lab demonstrations have historically run years ahead of qualified product — and whether Western Digital's bet on stretching conventional recording proves cheaper than Seagate's earlier move to lasers. Even the 1956 figures remain unsettled: IBM and the Computer History Museum still disagree on the character encoding of the first hard drive ever shipped.

Sources

IBM Heritage — RAMAC (primary) — lab history, Johnson quote, character encoding, early customers, cost comparison
Computer History Museum, The Storage Engine — 1956: First commercial hard disk drive shipped (primary / archival) — 3.75 MB, 50 discs, 1,200 rpm, 2,000 bits/in², 800-microinch flying height, Zellerbach Paper, $750 monthly lease, contemporaneous patents and papers
Seagate Technology — Seagate Delivers Industry's Highest Capacity Hard Drives with Next-Generation Mozaic 4+ (March 3, 2026) (primary) — 44 TB qualification, 4+ TB per disc, 10 TB per disc and 100 TB roadmap, one-exabyte efficiency figures
Western Digital — Western Digital Accelerates Storage Innovation for AI Era (February 3, 2026) (primary) — 40 TB UltraSMR qualification, HAMR ramp in 2027, ePMR extension
IBM Heritage — Reynold B. Johnson (primary) — 1952 West Coast laboratory assignment
Western Digital — Why Cloud Storage Runs on Hard Drives, Not SSDs — share of cloud exabytes stored on HDD (manufacturer analysis)
Tom's Hardware — Western Digital details 14-platter 3.5-inch HAMR HDD designs with 140 TB and beyond — longer-range platter roadmap
EDN — IBM intros 1st computer disk storage unit, September 13, 1956 — announcement date

Prices and specifications verified September 12, 2026 and subject to change. Areal density and cost-per-terabyte calculations marked as estimates are GadgetGlow Bytes arithmetic derived from the sourced figures above, not manufacturer specifications. GadgetGlow Bytes does not test hardware and does not receive products from manufacturers for coverage.

Search This Blog

GADGETGLOW BYTES