AI & Finance Briefing
9 min read

Nebius Priced $5B of Notes as SK hynix Set a ₩40T Buyback

Nebius upsized a convertible issue to $5 billion as SK hynix authorised a ₩40 trillion buyback, exposing opposite capital flows across the AI stack.

A monumental aged-brass screw jack lifts an unfinished cream paper data-centre roof from a rolled blank bond.
AI & Finance Briefing / 9 min read
AIENGINE

9 min read

Share

The AI capital cycle moved in two directions over the last 24 hours. Nebius priced $5.0 billion of convertible senior notes to finance data centres, GPUs and its full-stack cloud. SK hynix, one layer deeper in the supply chain, authorised an estimated ₩40.0 trillion open-market buyback and plans to cancel the shares.

Those are not contradictory signals. They show where cash is being consumed and generated. An AI cloud operator is using cheap-looking coupons, future accretion and potential equity issuance to bring capacity forward. A memory supplier is directing free cash flow back to shareholders. Kingsoft Cloud sat between them: AI cloud billings grew strongly and quarterly operating profit turned positive, but capital expenditure including leased assets exceeded revenue. BILL showed a software version of the same discipline, pairing AI-adoption language with restructuring, a smaller reported solution count and a large buyback.

In brief, for the window from 19 August 09:02 to 20 August 09:02 in Tehran:

  • Nebius upsized its proposed convertible financing from $4.5 billion to $5.0 billion and expects about $4.94 billion of net proceeds before any purchaser option.
  • The two Nebius tranches carry 0.50% and 4.50% cash coupons, but their principal accretes to 110% and 125% by maturity, so coupon alone understates the financing cost.
  • SK hynix approved an estimated ₩40.0 trillion purchase of 24.07 million shares, followed by cancellation, as part of a policy to return more than half of 2025–2027 cumulative free cash flow.
  • Kingsoft Cloud reported its first positive GAAP operating margin while spending RMB3.3 billion on capacity, including assets acquired through leases.
  • BILL reported 16% core-revenue growth, yet its GAAP operating loss widened after $76.6 million of restructuring expense and its reported count of businesses using solutions fell sequentially.

The evidence ledger

DisclosurePublished or filed in the windowConfirmed factImportant limit
Nebius pricing19 August; page updated 20 August 03:15 UTC$5.0bn base issue; about $4.94bn expected net proceedsSettlement is expected on 24 August and remains conditional
SK hynix capital return19 August; SEC filings accepted from 07:14 EDTEstimated ₩40.0tn buyback and cancellation planShare count and actual spend move with the purchase price
Kingsoft Cloud Q219 August; SEC filing accepted 06:00 EDTRMB23.0m GAAP operating profit; RMB3.3bn capexResults are unaudited and AI revenue is reported through company-defined measures
BILL fiscal Q419 August; SEC filing accepted 16:02 EDT$436.2m revenue; $34.3m GAAP operating lossThe release gives no AI-specific revenue or usage measure

Nebius priced leverage, not free capital

Nebius's pricing notice dated 19 August divides the issue into $3.0 billion of 0.50% notes due 2030 and $2.0 billion of 4.50% notes due 2034. The issue was upsized by $500 million from the proposal filed earlier that day. Purchasers also have options for another $750 million. If fully exercised, Nebius estimates net proceeds of approximately $5.68 billion rather than $4.94 billion.

The visible annual cash coupon on the base issue is $105 million: $15 million on the 2030 tranche and $90 million on the 2034 tranche. That is not the complete cost. Principal accretes to $3.3 billion and $2.5 billion respectively at maturity. If the notes remain outstanding and are not converted or redeemed earlier, their combined accreted principal would therefore reach $5.8 billion.

The embedded equity option matters too. Initial conversion prices are approximately $313.46 and $324.65 per share, 40% and 45% above Nebius's stated 19 August closing price. Applying the disclosed conversion rates to the base principal gives about 15.7 million underlying shares; this is an AIEngine calculation, not a forecast of issuance. Nebius may settle conversions in cash, shares or a combination, subject to the terms.

Separately, Nebius agreed to exchange $800 million of older notes for approximately 15.8 million shares. That removes old principal but creates near-term equity supply. The company warns that participating holders may sell shares or adjust hedges. Calling the whole transaction either “debt” or “dilution” misses its moving parts: cash coupons, accretion, conversion conditions, exchange shares and settlement choice each affect a different ledger.

The financing closes one gap and opens another

Nebius says proceeds will fund data-centre construction, footprint expansion, GPUs and its AI cloud. That use is consistent with its 12 August [financial review](https://www.sec.gov/Archives/edgar/data/1513845/000110465926094844/nbis-20260812xex99d1.htm), which reported about $5.7 billion of second-quarter capital expenditure and $8.0 billion of cash at June-end. The same quarter produced $2.25 billion of operating cash from continuing operations, much of it supported by customer prepayments.

Our earlier analysis of Nebius's prepayment-backed capacity separated contracted power, commissioned hardware, customer cash and recognised revenue. The new notes add a fifth record: financing that must eventually be repaid, refinanced or converted. It can fund the physical gap between an order and a live cluster, but it does not prove that power, buildings, GPUs and customer acceptance will arrive together.

An operator assessing a comparable financing should reconcile:

  • gross issue size to net cash after discounts and expenses;
  • cash coupon to accretion and other non-cash financing cost;
  • existing debt retired to new debt and shares issued;
  • funded construction to contracted, energised and commissioned capacity;
  • customer prepayments to remaining capital at risk; and
  • conversion scenarios to cash repayment and future share count.

The AI production cost-stack guide applies the same principle at workload level: cheap-looking compute is useful only when financing, utilisation, energy, operations and replacement are kept in the calculation.

SK hynix sent AI cash in the opposite direction

SK hynix's 19 August acquisition filing authorises open-market purchases from 20 August through 19 November. The company estimated 24.07 million shares and ₩40.004 trillion using the previous day's ₩1.662 million closing price. That estimated share count is about 3.3% of the 730.49 million issued shares disclosed in its matching cancellation decision.

The cancellation is substantive: issued shares will fall, although stated capital will not change under the structure described. The exact number is not fixed because open-market prices will determine how many shares ₩40 trillion buys. The company plans to cancel all acquired shares in one batch after purchasing finishes; it did not give the cancellation date.

Its separate shareholder-return policy says more than 50% of cumulative free cash flow for 2025–2027 will be returned through buybacks, cancellations and dividends. This is the other side of the AI infrastructure cycle. Memory demand can generate distributable cash at a component supplier even while cloud customers and operators raise new capital to buy the hardware.

That does not establish a permanent margin or demand level. It establishes a capital-allocation decision made from current free cash flow and management's view that the shares are undervalued. Procurement teams should not read the buyback as a capacity commitment; investors should not read an estimated purchase value as money already spent.

Kingsoft Cloud crossed operating breakeven, but not the funding finish line

Kingsoft Cloud's 19 August unaudited results reported AI cloud gross billings of RMB1.327 billion, up 82% year on year and equal to 56% of public-cloud revenue. Total revenue rose 30.8% to RMB3.072 billion. GAAP operating profit reached RMB23.0 million, a 0.7% margin, for the company's first positive quarter on that measure.

The cash and asset bridge remains demanding. Capital expenditure, including capitalised assets through leasing, was RMB3.3 billion—more than quarterly revenue. Depreciation and amortisation in cost of revenue rose 74.6% to RMB963.8 million. Cash and equivalents fell from RMB6.018 billion at December-end to RMB4.674 billion at June-end, although second-quarter operating cash flow was positive.

AI growth was not the only reason operating profit improved. Operating expenses fell 33.4%, including lower credit-loss expense, personnel cost and share-based compensation. The company still recorded a RMB93.0 million net loss after interest and other items. The operational reading is progress, not completion: AI demand is contributing more gross profit, while server acquisition, leasing, depreciation and finance still determine the return.

BILL's AI claim needs its own revenue line

BILL's 19 August fiscal-year results reported $436.2 million of quarterly revenue and 16% core-revenue growth. Management said adoption of its AI capabilities was increasing, but the release does not quantify AI users, AI revenue, task volume, accuracy or retention.

The disclosed operating changes are more measurable. GAAP operating loss widened to $34.3 million from $22.3 million after $76.6 million of restructuring expense. Research and development expense fell to $63.1 million from $90.1 million. The reported count of businesses using solutions fell to 479,300 from 493,800 in the preceding quarter; the metric is not a unique-customer count because businesses using multiple products may be included more than once. BILL also spent approximately $300 million repurchasing shares during the quarter.

None of this disproves productive AI adoption. It sets the evidence requirement. A finance platform should eventually connect AI usage to verified task completion, payment or expense volume, retention, gross profit and support cost. Without that bridge, “AI adoption” remains management commentary beside accounts driven by the wider platform.

Limits and what to watch next

Nebius had priced but not settled the notes by the cutoff. Its accretion, conversion and redemption mechanics are complex, and actual cash or share settlement depends on future conditions. SK hynix's purchase quantity changes with its share price. Kingsoft Cloud and BILL reported unaudited quarterly figures, and each company defines non-GAAP and operating measures differently. Currency translations and cross-company margin comparisons can mislead. This brief describes financing and operating evidence, not investment advice.

The next proof points are concrete:

  • Nebius's 24 August settlement, final option exercise, exchange completion and resulting cash and share count;
  • construction spend reconciled to commissioned megawatts, customer acceptance and recognised cloud revenue;
  • SK hynix's actual average purchase price, acquired share count and final cancellation disclosure;
  • Kingsoft Cloud's capex, lease liabilities, depreciation and cash conversion after the first positive operating quarter;
  • BILL's post-restructuring product investment and a quantified link between AI use and durable customer economics; and
  • whether other profitable AI suppliers return cash while capacity operators continue to issue equity-linked finance.

The day's message is not that one end of the AI stack is healthy and another is weak. It is that the same demand creates different capital obligations at each layer. The useful control is to follow every dollar from security issuance or operating cash through construction and service delivery, then ask whether it returns as durable free cash flow before the financing claim comes due.

Primary sources

TaggedAI InfrastructureConvertible NotesNebiusSK hynixKingsoft CloudBILL
Work With Us

Interested in implementing this for your business?

We help UK businesses put these ideas into practice. Book a call to discuss your specific situation.