AI Infrastructure Financing Digest October 22 2026 Shows Short Debt Funding Long Bets
Six financing announcements crossed the wire on October 22. None of them is a venture round. That’s the point. The AI build-out has moved past the stage where equity pays for it, and the money now comes as bank lines, bridge loans, equipment leases and government checks with strings attached. Read the headline number on each and you learn very little. Read the tenor, the rate and who carries the risk, and the picture gets clearer.
Nebius signed a $3.5 billion senior secured facility with Citi, which acts as lead arranger and agent. The money goes to data centers, GPUs and general corporate needs. The detail that matters is the term: 364 days. Pricing starts at SOFR plus 2.75 points and can climb to 3.50. Nebius points to more than $40 billion of long-term contracts with investment-grade customers as the backing, and it’s chasing a 4 GW capacity target with a gigawatt-scale campus in Missouri already under way. So the assets last years and the contracts run for years, but this loan has to be repaid or rolled within twelve months. It sits on top of roughly $775 million of GPU-backed debt raised in July and $5.75 billion of convertible notes from August. A one-year facility for a multi-year build is a bet that credit markets stay open next October.
Atlas Energy Solutions is a Permian frac sand and oilfield logistics company. On October 22 it closed a $424 million delayed-draw bridge loan arranged by Goldman Sachs and Barclays, with $171 million drawn at closing. The money pays for power generation equipment at two data center projects due to start operating in late 2027. The initial maturity is 180 days, extendable by another 184 if lenders agree and Atlas signs energy supply agreements. The rate is SOFR plus 2.50, stepping up half a point after six months, and Atlas can ask for up to $286 million more. Read that closely. A sand company is buying turbines before the customer contracts that justify them are signed, and the loan’s extension depends on those contracts showing up. The company is also running operating losses.
Helion and Vertical Semiconductor were two of four companies in a Commerce Department round of CHIPS letters of intent totalling up to $389 million. Commonwealth Fusion Systems got up to $85 million and quartz supplier Sibelco up to $56 million. The condition is new: Washington takes a minority, non-voting equity stake in every company it funds.
Helion’s share is up to $175 million over three years. More than $100 million goes to high-voltage semiconductor and packaging work in Everett, Washington, and about $80 million to chip development with GlobalFoundries. Helion is adding $50 million of its own. Fusion needs these parts because Helion’s machine pulses huge currents through magnets and recovers electricity straight from the expanding plasma, with no steam turbine in between. Switches that can survive that load are mostly bought abroad today. Orion, the plant meant to sell power to Microsoft, is under construction in Malaga, Washington, with first operations targeted for 2028.
Vertical Semiconductor, an MIT spin-out built on ARPA-E-funded work, gets up to $73 million for vertical gallium nitride power transistors. The target is a 2.5 MW AI rack that the company says cuts energy losses by 30% and footprint by half. Its CEO calls power the biggest bottleneck in AI infrastructure. That’s hard to argue with after reading about Atlas buying turbines.
All of these are letters of intent, not final awards. The equity stake is the part that will outlast this round.
QumulusAI, an Atlanta GPU cloud focused on inference, expanded its equipment lease financing with Tech Finance Co. to about $123.6 million, up from $26 million in May. The leases fund NVIDIA Blackwell GPUs bought in May and July. TFC covers 85% to 90% of each equipment schedule, customer deposits cover the rest, and the leases run three years at an implied high single-digit rate, with a buyout or return option at the end. “No new equity” is the selling line. The less advertised point is that a three-year lease on a chip generation that may be superseded within two years leaves somebody holding residual value risk, and the end-of-term option shows it’s negotiable.
SOAR AI, the NYSE American company that renamed itself on October 19 after merging with Alignment Engine, published an investor deck around $1.17 billion of contracted orders. Those orders are a master services agreement and two four-year take-or-pay contracts with a single unnamed AI customer, which SOAR itself flags as a concentration risk. The first tranche is 4,096 AMD MI355X GPUs at a North Canton, Ohio campus, due in service by December 31 with internal power-on November 30 and a three-year revenue backstop of about $215 million. A second tranche of 4,032 MI455X GPUs could arrive as early as April 2027. The site has 154 MW today and a roadmap to 480 MW. SOAR guides to $175 million to $200 million of 2027 revenue and about $294 million a year at full run-rate. Each phase only starts once its contracts and financing are in place, and that financing is still prepayments, equipment debt, senior debt and equity, in that order.
Put the six side by side and the pattern is easy to see. Loans of 180 and 364 days, three-year leases, a government that wants equity, a sand company buying power plants and a one-customer order book. The contracts these structures lean on are real. Every deal still assumes the next round of money arrives on time.
That assumption is where to watch.