In partnership with
Why would a company turn down a $1.6bn buyout?
To raise two thirds as much at a 7x higher valuation, and open up another front in the fight against the outfit that sells almost every AI chip on earth. Not with a breakthrough, but with conviction.
And a great deal of other people's money.
Because in this market, that’s just as good.
I'm Ben Baldieri. Every week I break down the moves shaping GPU compute, AI infrastructure, and the data centres that power it all.
Here's what's inside this week:
Let's get into it.
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
SambaNova Raises $1B at $11B and Lands JPMorgan
One doesn’t turn down a buyout offer from one of the largest semiconductor companies on the planet without good reason.
SambaNova Systems raised $1bn at an $11bn valuation, the first close of its Series F, led by General Atlantic. Seligman, T. Rowe Price and Capital Group joined, and a second close is expected within weeks. The round lands five months after the company's $350m Series E and the February launch of its SN50 inference chip. It also comes after SambaNova walked away from Intel acquisition talks that valued it near $1.6bn, reported by Bloomberg in December. The company builds reconfigurable dataflow silicon for inference, and CEO Rodrigo Liang has framed a public listing as the likelier path than a sale.
Why this matters:
Alt-silicon is drawing real conviction capital. General Atlantic led at $11bn, roughly seven times what Intel was reportedly willing to pay in December, a mark on how fast the anti-Nvidia trade has repriced.
It rhymes with last week. In Issue #113, Nvidia sat on the cap tables of both Firmus and Together AI; in Issue #112, OpenAI taped out its own inference chip with Broadcom. SambaNova is the pure-play counter: a merchant inference challenger with no Nvidia on the register.
The customer question just got a name. JPMorgan is deploying SambaNova's SN40L and SN50 for on-prem inference, the proof point an $11bn mark needs. Watch whether the second close brings more banks like it, or just more capital.
SK Hynix Raises $26.5B in Its US Debut
The biggest US listing ever by a foreign company just priced, and it belongs to a memory maker, not an AI lab.
SK Hynix priced its US debut at $26.5bn, selling 177.9m American depositary shares at $149 each on the Nasdaq under the ticker SKHY. The Korean group is the world's largest maker of high-bandwidth memory, the stacked DRAM beside every AI accelerator, and its first-quarter revenue grew about 198% year on year per its SEC filing. Demand ran more than seven times the shares on offer, and the ADRs opened 14% above the $149 price on 10 July. Proceeds go to capex: new fabs in South Korea and the EUV lithography scanners used to make advanced chips. SK Hynix's market value crossed $1tn in May, the same month Micron did.
Why this matters:
At $26.5bn it is the largest US listing ever by a foreign company, past Alibaba's 2014 debut, and second only to SpaceX's $85.7bn June IPO among all US listings. Public markets, not private capital, are underwriting the AI memory cycle now.
SK Hynix, Micron and Samsung are the entire high-bandwidth-memory supply base, and two of the three are raising to expand it this quarter. Micron's revenue quadrupled to $41.45bn in Issue #112, and the same shortage is what SK Hynix just sold to Nasdaq.
Listing in New York rather than topping up its Seoul float aims US public money straight at the AI supply chain. Expect more of that chain, not just the model labs, to chase the same bid.
Twelve Banks Hand Nscale a $900M Revolver
A neocloud dogged by a lost lawsuit and loan-default questions just pulled a $900m credit line from twelve of the world's biggest banks.
Nscale closed a $900m revolving credit facility syndicated across J.P. Morgan, Goldman Sachs, Morgan Stanley, MUFG, RBC, Bank of America, Crédit Agricole, Deutsche Bank, Mizuho, SMBC, TD and KeyBank. A revolver lets it draw, repay and re-borrow against build-out across the US, Europe and APAC. The Nvidia-, Dell- and Nokia-backed provider was valued at $14.6bn after a $2bn round in March, and already sells capacity to Microsoft and OpenAI. A day earlier it signed a partnership with Nordkraft to support data centre operations in Narvik, Norway.
Why this matters:
Revolver access is a graduation. Syndicated bank credit is cheaper than the venture equity and project debt neoclouds have leaned on, and it means lenders now treat GPU build-out as investment-grade-adjacent, the shift behind IREN's investment-grade GPU debt in Issue #109.
Neocloud financing is splitting by instrument. Together AI raised $800m of equity last week; Nscale drew $900m of credit this week. Same build-out, different paper, and the gap in cost of capital between tiers is the whole game.
Meta Ships Muse Spark 1.1, OpenAI Makes GPT-5.6 Microsoft's Default
Two frontier models shipped on Thursday, and the more telling one came from the lab we thought was leaving the race.
Meta released Muse Spark 1.1 from its Superintelligence Labs, a multimodal reasoning model built for agentic coding and tool use. It put the model behind its first paid API, at $1.25 per million input tokens and $4.25 per million output, aimed straight at the OpenAI and Anthropic coding market.
The same day, OpenAI shipped GPT-5.6 and made it the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat and Cowork. That was the same week Bloomberg reported Microsoft was leaning on its own MAI models to cut costs in those apps.
Why this matters:
Meta is not leaving the frontier race; it is monetising it. In Issue #113 we asked whether Meta was quietly dropping out. Muse Spark 1.1, its first model sold by the token, is the answer.
GPT-5.6 as the Microsoft 365 default reads as a defensive flex. You announce a partnership loudest when the reporting says it is fraying.
Both launches feed the same machine. Cheaper, more capable models pull more inference through the data centres the rest of this issue is busy financing, and that demand is not slowing: last week Google was rationing Meta's Gemini access for lack of capacity (Issue #113).
Micron Adds $3B to a $250B US Memory Bet
The real GPU bottleneck is the memory stacked next to it, and Micron just poured concrete on the answer.
Micron committed up to $3bn to strengthen the US chip supply chain, including $500m for GlobalWafers to expand in Texas. It also poured first concrete at its Clay, New York megafab, more than a quarter ahead of plan. That build sits inside a US programme Micron now puts above $250bn through 2035, aimed at making 40% of its DRAM at home. Micron also signed a long-term memory-supply deal with Ford. High-bandwidth memory is the stack of DRAM beside each accelerator, and demand for it has largely outpaced supply through the current cycle.
Why this matters:
HBM is made by only three firms at scale: SK Hynix, Samsung and Micron. Pulling a slice of that supply onshore is a strategic move, not just a commercial one.
A $250bn forecast is Micron reading the demand curve. You do not break ground on a four-fab campus unless you expect memory to stay short, and Micron's revenue quadrupling to $41.45bn (Issue #112) says it has been.
It is the supply-side answer to the efficiency wave. Frameworks like DeepSeek's DSpark from Issue #113 squeeze more out of each chip; Micron is betting demand still outruns the squeeze.
Amazon Borrows $25B, Some of It Until 2066, for AI
Amazon put a price on the AI build-out, and it is 6.25% at the long end.
Amazon raised up to $25bn across eight tranches, from floating-rate notes due 2029 to fixed-rate paper due 2066, with coupons between 4.6% and 6.25%. Order books hit $62bn before settling near $41bn, with Barclays, Goldman Sachs, JPMorgan and Morgan Stanley running the sale. Proceeds go to "general corporate purposes." Amazon expects the bulk of a roughly $200bn 2026 capex programme to land on AWS AI and data centre build-out.
Why this matters:
The hyperscalers are funding AI on the bond market now, not just out of cash flow. A $25bn raise inside a $200bn capex year is the balance sheet stretching to meet the model roadmap.
The maturities tell the story. Forty-year money against chips that depreciate in a handful of years is a bet that the demand outlives any specific GPU.
It extends the debt thread from Issue #113, where SoftBank revived a $10bn loan against its OpenAI stake, and the $120bn of new Meta and Microsoft lease commitments in Issue #112. The cheapest capital in the market is still the giants', and they are drawing on it hard.
One Virginia Data Centre Just Backed $520M in Bonds
One data hall's rent got turned into a bond this week.
Cloud Capital, through a vehicle called Lohrasp Enterprise II, issued $520m of commercial mortgage-backed securities against a single data centre in Loudoun County, Virginia. The 80MW facility, LC3, finished construction this year and is fully leased to one investment-grade hyperscale tenant on a long-term triple-net lease that throws off about $65.1m in annual base rent. The notes are interest-only, with a five-year anticipated repayment and a 30-year legal maturity. One building, one tenant, one rent cheque, tranched and sold to the bond market.
Why this matters:
Data centre rent is collateral now. It is the same shift that took IREN's GPU debt investment-grade (Issue #109) and let Aligned tranche $1.19bn of rent into bonds last week (Issue #113): package the lease, sell the paper, fund the build-out below the cost of equity.
The single-asset, single-tenant structure concentrates the risk. If that one investment-grade hyperscaler ever walks, there is no diversification under the notes; the cash trap fires and equity takes the first loss.
It is the week's cleanest expression of the whole issue. The data centre is no longer just the asset, it is the security, and the capital stack now runs from the GPU all the way out to a 30-year bond.