The Refused Test Is the Finding

The Core Insight

Theranos raised 900 million dollars and reached a valuation of 10 billion dollars. Its own devices ran 12 of the 250 tests on its menu.

John Carreyrou broke the story at the Wall Street Journal and published Bad Blood in 2018. The SEC charged Holmes, Balwani and the company in March 2018 with an elaborate, years-long fraud. Holmes gave up voting control, paid a penalty of 500,000 dollars, and took a ten-year officer and director bar. Read this as reporting on documented fraud, not as a study of a business judgment that went wrong.

Most founders read the story as one unusual liar, and conclude that judgment about people is the defense. Carreyrou's account shows six control surfaces failing at once: the board, finance, the regulator, the retail partner, the press, and the employees. Character reading is the weakest item on that list.

Each of those surfaces had a cheap test available. The list is short: a fifty patient comparison study, a split sample sent to Quest, one device shipped to an outside lab. Someone asked for each one, and each was refused.

The Framework

The thesis is structural. A company holds a fifteen year gap between claim and capability when every mechanism that closes the gap is disabled. Theranos disabled six.

• The board carried prestige and no blood science, so no director evaluated an assay.
• Holmes took one hundred votes per share in December 2013, or 99.7 percent of the vote.
• The company ran without a CFO after 2006, so contradictory projections lived in one quarter.
• Engineering and chemistry reported only to Holmes, so no whole system test ever ran.
• Regulatory jurisdiction split between the FDA and CMS, and California underfunded its lab inspectors.
• Counsel took stock instead of fees and sat at every board meeting from 2013.

The capability side of the gap ran through three generations. The 2004 microfluidic cartridge and reader never reproduced a result from the same sample. Tony Nugent replaced it in 2007 with a glue-dispensing robot bought for 3,000 dollars, which employees called the gluebot. The miniLab that followed had to cram a spectrophotometer, a cytometer, an isothermal amplifier and a photomultiplier tube into one box. Its throughput was one sample at a time.

None of those instruments was new. An off-the-shelf competitor, the Piccolo Xpress, already ran 31 tests in as little as 12 minutes on three or four drops of blood.

The money and the partners tracked the story, and the hardware never caught up. Theranos raised nearly 6 million dollars in 2004. Partner Fund Management put in 96 million dollars in 2014, at a valuation of 9 billion dollars. Walgreens paid an innovation fee of 100 million dollars and lent 40 million dollars more. All of it was sunk before Walgreens terminated the deal on June 12, 2016. Safeway loaned 30 million dollars and spent 350 million dollars renovating more than half of its 1,700 stores.

Key Ideas

The Device Ran 1980s Chemistry on a Glue Robot

Nugent fastened a pipette to the robot and programmed it to copy a bench chemist. It ran chemiluminescent immunoassays, a technique pioneered in the early 1980s at Cardiff University. It needed about 50 microliters of blood, five times the 10 microliter figure Holmes promised.

The limit was categorical. The Edison ran immunoassays and nothing else. General chemistry, hematology and DNA amplification sat outside its reach, and the chem 18 panel accounted for about two-thirds of doctors' orders.

Its pipette tips were imprecise, so the protocol compensated by brute force. Dilute the sample, split it three ways, run it on three Edisons with two tips each, and report the median of six values. The device also had to hold 34 degrees Celsius, maintained by two 11-volt heaters that failed in cold rooms.

Holmes set the 10 microliter specification herself, and no physics required it. That spec forced heavy saline dilution, which degraded every measurement downstream.

The Patient Results Came Off Hacked Siemens Machines

By mid-2013 Holmes had a launch deadline and no working device. Daniel Young and Sam Gong were sent to hack the Siemens ADVIA 1800, a commercial analyzer weighing 1,320 pounds. They used a Tecan liquid handler to pre-dilute finger-stick blood with saline, in cups half the normal size, to defeat the dead volume.

Three consequences were known inside the company. The ADVIA dilutes internally, so the blood was diluted twice. Double dilution pushed concentrations below the FDA-sanctioned measurement range, and the reported value came from multiplying back up. The third was off-label use of a third party's regulated instrument, with Alan Beam's name on the CLIA certificate as lab director.

Theranos bought six more ADVIAs and put up partitions so Siemens service technicians did not see the Edisons. Of more than 240 tests on the menu, about 80 used finger-stick blood. Roughly a dozen ran on an Edison, and 60 or 70 on hacked Siemens machines. Everything else was a conventional venous draw.

The validation data was built to match. A syphilis study of 247 samples with 66 known positives detected 65 percent on the first run and 80 percent on the second. The report claimed 95 percent sensitivity, reached by widening an equivocal zone with no predefined criteria.

The November 2015 CMS report ran 121 pages. Edisons failed quality control nearly a third of the time in one month, and one hormone test failed 87 percent of the time. CMS declared immediate jeopardy to patient health and safety. The company voided or corrected nearly 1 million test results.

Governance Was Engineered to Be Uncheckable

The late board was George Shultz, Henry Kissinger, William Perry, Sam Nunn, Gary Roughead, James Mattis, Richard Kovacevich and Bill Frist. The common denominator was the Hoover Institution, and seats were offered in exchange for grants of stock. Frist, a transplant surgeon, was the only physician. No director had knowledge of blood science, and no health-care venture firm ever invested.

In December 2013 Holmes forced through a resolution assigning one hundred votes to every share she owned. That gave her 99.7 percent of the voting rights, and the board did not reach a quorum without her. Shultz, deposed later, said they never took any votes at Theranos. No independent investigation was ever commissioned.

The finance gap is measurable. Balwani sent Partner Fund projections of 165 million dollars gross profit on 261 million dollars of revenue for 2014. Six weeks later the controller sent an options-pricing firm 35 million dollars profit on 50 million dollars of revenue for the same year.

In March 2008 the sales head and the general counsel told the chairman that the projections were irreconcilable with the product. Four directors resolved to remove Holmes as chief executive. She talked them out of it in two hours, and both men were fired within weeks.

Compartmentalization Blocked the Internal Correction

Compartmentalization was designed in from the first months. Engineering and chemistry both reported to Holmes and were not encouraged to talk to each other, so only she held the full picture. Aaron Moore observed that each group tested its own parts and nobody ran overall system tests.

Theranos blocked chat ports, reviewed badge logs each morning, and required an NDA from anyone who entered the building.

The physical expression was a two room lab. Jurassic Park held the commercial analyzers, carried the CLIA certification, and was what inspectors saw. Normandy held the Edisons and the hacked Siemens machines behind a key-carded door. During the December 2013 state inspection, staff were ordered to stay out of Normandy, and the inspector never saw it.

The Departures Were the Highest Fidelity Signal

The treatment of dissenters is consistent across a decade. Henry Mosley objected to faked investor demos in 2006 and was fired that day. Avie Tevanian documented irreconcilable discrepancies as a director in 2007 and was pushed to resign. The general counsel sent him a Christmas Eve email threatening suit for breach of fiduciary duty.

Ian Gibbons was demoted in 2013 after his doubts reached Holmes, then subpoenaed in a patent case. A company lawyer emailed him a doctor's note template to help him avoid the deposition. He took a fatal overdose hours later. No memorial service was held, and a Theranos lawyer emailed the same day demanding his laptop and phone.

Tyler Shultz sent his concerns by email in 2014 and drew a rebuttal from Balwani demanding an apology. Two Boies Schiller partners ambushed him at his grandfather's house with a temporary restraining order. He was pressured across four drafts of an affidavit to name the Journal's other sources. His parents spent more than 400,000 dollars, and he never signed.

Alan Beam deleted 175 forwarded emails on the advice of his own intimidated lawyer. Erika Cheung was served at a home address that only surveillance produced. Mosley, Tevanian, Dupuy, Laghari, Beam, Shultz and Cheung were right, in order, for nine consecutive years.

Every Detection Cost Less Than a Week

The demos failed whenever an outsider set the terms. One of two readers malfunctioned at Novartis in November 2006, and the team in California beamed over a fabricated result. Investor demos displayed a pre-recorded result from an earlier successful run. Balwani later had an engineer write software that suppressed miniLab error messages and showed a crawling progress ring.

MedVenture Associates asked in July 2004 how the microchip differed from the commercial product Abaxis already sold. Holmes had no answer and left after about an hour.

Kevin Hunter, the consultant Walgreens paid 25,000 dollars a month, asked for a fifty patient comparison against Stanford Hospital. He was denied that, an embedded observer, and access to the lab. He drew the blood of the Walgreens pharmacy president and never got a result back. Renaat Van den Hooff answered those red flags with the risk of CVS getting there first.

Kent Bradley at Safeway retested discrepant employees at Quest and LabCorp in 2012 and built a discrepancy analysis. A false elevated PSA reading on a senior executive came back normal elsewhere. He escalated, and Steve Burd brushed him off.

Adam Clapper read the one peer-reviewed citation Holmes offered in December 2014. The journal charged a publication fee of 500 dollars, and the paper covered one blood test on six patients. That check took twenty minutes.

Any physician after September 2013 had the cheapest test of all. The cortisol of Doctor Sundene read under 1 microgram per deciliter at Theranos and 18.8 at LabCorp.

Practical Applications

Write down the cheapest experiment that shows the central claim false, then ask for it. The menu is short: a split sample, a device shipped to a lab you pick, an observer in the room for one week. A refusal is the result.

Read the control structure before the resumes. Ask what fraction of the votes the founder holds, when that fraction was set, and whether any director understands the core technical claim.

Ask for both sets of numbers. Compare what investors are shown against what the option pricing and the payroll assume. A gap of five times inside the same quarter is the most mechanical warning sign in the book.

Read a third party validation before repeating it. The Johns Hopkins validation Theranos cited was a two-page meeting summary carrying an explicit non-endorsement disclaimer. Theranos still told Walgreens that ten of the fifteen largest pharma companies validated its system comprehensively.

Track who leaves, and weigh it above what the company announces. Count the senior people who left inside two years, and ask which of them raised a technical objection first.

Who This Is For

Founders raising money, and anyone who sits on a board or writes a check, get the most from this. It is also the clearest case study available for anyone building in a regulated field, where a wrong answer lands on a patient.

Hindsight makes these signals look sharper than they were on the day. Carreyrou wrote after the collapse, with subpoenaed documents and more than sixty former employees willing to talk, while Holmes declined to cooperate. Many investors, partners and reporters held the same information in real time and did nothing. Take the lesson as a case for verification systems rather than for reading character. The people who missed it were experienced and well informed.

Skip it as a management book. This is reporting, and the transferable content sits in the control failures.

The Decision

Pick one claim you currently believe because a credible person repeated it. That covers a vendor accuracy figure, a partner forecast, or a number in your own deck that nobody has traced to raw data.

Name the cheapest experiment that shows the claim false, and cap it at one week. Then ask the person who holds the claim to run it with you watching.

The answer to that request carries more information than the answer to the original question. Record what you get.