The venture industry's operating system
SVB was conceived in 1982 by Bill Biggerstaff, a Wells Fargo executive, and Robert Medearis, a Stanford professor, and opened its first office in San Jose in 1983. Its entire commercial idea was to bank the customers other banks found unbankable: pre-revenue technology companies with no collateral, no earnings and a business plan written in the future tense. This worked extremely well. By the mid-2010s it was serving close to two-thirds of all US startups, and by its 2021 peak it had banking relationships with nearly half of all US venture-backed startups.
That is a triumph of positioning and, simultaneously, the whole problem. A bank whose depositors are all in the same industry, in the same few zip codes, funded by the same few dozen venture firms, and connected by the same group chats does not have a diversified deposit base. It has one very large depositor wearing a thousand hats. The balances were also enormous by retail standards, which meant most of them sat far above the FDIC insurance limit and were therefore genuinely at risk in a failure — the precise condition under which running early is the rational thing to do.
HTM accounting and the
During the 2020–21 tech boom, deposits poured in faster than SVB could lend them out, and the bank put a large share of the money into long-dated Treasuries and mortgage-backed securities bought at the very low yields available at the time. Then the Federal Reserve raised rates, and the market value of those fixed-rate bonds fell, because that is what bonds do. See for why the long ones fell hardest.
What kept this out of the headline numbers was an accounting distinction. Securities classified as available-for-sale are to market, so unrealised gains and losses show up in equity and are visible in reported book value. Securities classified as held-to-maturity are carried at amortised cost, and their unrealised losses generally do not run through earnings or capital at all, provided the bank asserts that it intends and is able to hold them to maturity. SVB began shifting large volumes of securities from AFS into HTM in 2014, and by the end of 2022 held about $91.3bn of HTM securities, with roughly $15bn of unrealised losses sitting in the footnotes rather than in the income statement.
HTM accounting does not make a loss go away. It defers recognition of the loss for exactly as long as you never have to sell, and the one thing that forces a bank to sell is depositors asking for their money. This is the problem in its purest form: an asset can be worth two different numbers depending on whether anyone makes you find out. Short, flighty, uninsured funding against long, underwater assets is the oldest failure mode in banking, and no amount of classification fixes it.
The $42 billion bank run
On 8 March 2023, SVB announced that it had sold securities at a $1.8bn loss and intended to raise $2.25bn of new equity. The announcement was intended to reassure. It did the opposite, because it told a sophisticated customer base exactly two things: the bond book was underwater, and the bank needed money.
On 9 March, venture firms and founder networks amplified the alarm, including group-chat instructions to portfolio companies to their cash. Depositors withdrew about $42bn in a single day. There was no queue outside a branch. There was a wire instruction and a WhatsApp thread, and the run happened at the speed of a phone.
On 10 March, California regulators closed the bank and appointed the FDIC as receiver. On 12 March, the Treasury, the Federal Reserve and the FDIC announced that all SVB depositors would be made whole, invoking the systemic risk exception — the statutory override that lets regulators protect uninsured deposits when the alternative is contagion. Depositors had access to their money from Monday 13 March. Shareholders and management were not protected; the depositors above the insurance cap were, which is a policy choice with consequences that outlast the week.
The Federal Reserve published its own post-mortem in April 2023, led by Vice Chair for Supervision Michael Barr, examining both the bank's failures and the supervisory failures around it (the full report is here). Former chief executive Greg Becker testified before the Senate Banking Committee on 16 May 2023, alongside executives from Signature Bank, in a hearing at which he framed SVB as the victim of a confluence of factors. His written testimony is public. What has since happened to SVB's senior management by way of enforcement is, as far as we can establish, not much: recent coverage turns up supervisory reviews and regulatory argument rather than any resolved action against individual executives, and we are not going to assert an outcome the record does not show.
First Citizens, HSBC, and the non-bank shift
The carcass was distributed quickly. On 13 March 2023 the Bank of England sold Silicon Valley Bank UK Limited to HSBC UK Bank plc for £1; as at 10 March, SVB UK had loans of around £5.5bn and deposits of around £6.7bn. The Bank of England lists it among its past resolution actions, executed as a share transfer with a mandatory reduction of capital instruments. HSBC later rebranded the business as HSBC Innovation Banking.
On 27 March 2023, First Citizens BancShares bought $110.1bn of assets from the FDIC, including about $72bn of loans and $56.5bn of deposits, at a $16.45bn discount, with no cash paid upfront. The financing was a five-year, $35bn purchase money note issued to the FDIC, plus equity appreciation rights in First Citizens stock worth up to $500m, plus a loss-share arrangement under which the FDIC covers half of losses on the commercial loan portfolio above $5bn. First Citizens emerged with more than $218bn of assets and more than $144bn of deposits across more than 500 branches in 23 states. It was reported in July 2026 that First Citizens has repaid $8.5bn to the FDIC in connection with the acquisition.
The cost to the Deposit Insurance Fund is one of those numbers that gets quoted with false precision. One FDIC estimate of losses to the fund attributable to SVB alone is about $16.1bn. A separate and more recent figure — approximately $16.7bn as of 17 April 2026 — is the special assessment covering the failures of Silicon Valley Bank and Signature Bank combined, attributable solely to the guarantee of uninsured deposits. Those are not the same measurement, and anyone using them interchangeably is being sloppy.
The more interesting inheritance is the lending. First Citizens kept the SVB franchise running and HSBC took the UK book, but a meaningful share of venture lending migrated to specialist non-banks as bank lenders pulled back — firms such as Hercules Capital, TriplePoint and Runway Growth Capital. US venture debt reached $68.8bn in 2025, a record. So the business did not shrink. It moved to lenders that do not take deposits, are not examined by , and do not have access to the discount window. That is the same migration we track under , arriving in a different sector.
Why it matters to this crash
SVB is the reference implementation. Before March 2023 the standard modelling assumption was that a run takes days and involves physical queues; afterwards the assumption is that a concentrated, uninsured, well-connected deposit base can remove a fifth of a bank's funding between breakfast and dinner. Every stress test, deposit-beta model and liquidity buffer written since has had to account for that.
The underlying condition has improved but not disappeared. Unrealised losses on securities at US banks were about $325.1bn in the FDIC's Q1 2026 data, as of 31 March 2026 — roughly $375bn below the late-2022 to early-2023 peak of around $700bn, and about 46% of it. Which is to say: better, and still a third of a trillion dollars of losses that exist only as long as nobody is forced to sell.
The structure travels. We wrote on 19 August 2026 about Japan's life insurers, whose unrealised losses on domestic bonds reached ¥30.86tn ($194bn) at the end of June 2026, up 60% year on year, with Sony Life's lapse and surrender rate rising to 1.4% in April–June — the same arithmetic in a different jurisdiction. Long assets, hold-to-maturity comfort, and a liability side that can decide to leave. SVB is the case study for what happens when the third variable moves.
There is also a residue. The systemic risk exception, once used, is known to be usable. Depositors above the insurance cap at a mid-sized bank now hold something between a guarantee and a hunch, and the industry pays for it afterwards through assessments. That changes behaviour on both sides of the counter, and not obviously for the better.
What would make this dangerous
Start with the measurable things. First, the direction of the $325.1bn: if unrealised securities losses at US banks start climbing back toward the roughly $700bn peak because long yields rise again, the buffer that has been quietly rebuilding since 2023 goes into reverse. Second, concentration. Any bank whose HTM book is large relative to equity and whose deposits are dominated by uninsured balances from a single industry has the SVB configuration, regardless of the industry. The failure mode is not tech. It is correlation plus uninsured plus long bonds.
Third, the non-bank venture lenders. The $68.8bn US venture debt market of 2025 is a record set partly because specialist lenders absorbed share that banks vacated. Those lenders funded a cyclical borrower base without deposit insurance, discount-window access or the systemic risk exception standing behind them. A sharp tech funding downturn would test what happens when the lender of last resort is a fund's own credit line rather than the Federal Reserve, and nobody has run that experiment at this size.
Fourth, the First Citizens structure itself. The $35bn purchase money note and the loss-share arrangement covering half of commercial loan losses above $5bn are contracts with terms and dates. Repayments — $8.5bn as reported in July 2026 — are worth watching, because the deal's economics only look free while the acquired loan book performs.
And the unfalsifiable one that matters most: run speed. Nobody knows how fast the next one goes, and nobody had modelled $42bn in a day either.