Bloomberg Tech
2026-06-25 · Hosted by Caroline Hyde, Ed Ludlow · Bloomberg / iHeartMedia
Executive Summary
OpenAI unveiled “Jalapeno,” its first custom AI chip developed with Broadcom, claiming a 50% lower cost versus a typical AI GPU and targeting cheaper, faster inference — sending Broadcom shares up ~2% intraday. The memory trade dominated the conversation ahead of Micron earnings, with SK Hynix announcing a landmark $29.4 billion US ADR listing (top-five share sale of all time, trading July 10) to expand HBM capacity. Hyperscalers have now committed $850 billion in future data center leases led by Meta and Microsoft, while SpaceX completed a record $25 billion investment-grade bond sale and Cerebras reported its first post-IPO results — beating on revenue but falling ~17% on sequential margin contraction. Defense-tech and AI infrastructure financing remained hot themes, with Hadrian in talks to quadruple its valuation to $7.5 billion.
Key Stories & Changes
1. OpenAI’s “Jalapeno” Custom Chip with Broadcom
OpenAI unveiled the Jalapeno Intelligence Processor, a custom AI chip co-developed with Broadcom, claiming ~50% lower cost than a typical AI GPU
Focus is faster, cheaper AI inference; the move aims to diversify OpenAI away from heavy reliance on Nvidia and own more of the infrastructure stack
Broadcom CEO Hock Tan sees “every frontier lab” eventually building a custom chip, and predicts demand will exceed his prior 1.3-gigawatt supply estimate next year
OpenAI previously reported to spend tens of billions on Broadcom chips as part of hundreds of billions in infrastructure commitments; financing mechanism remains unclear
2. SK Hynix’s $29.4 Billion US Listing
SK Hynix plans to raise $29.4 billion in a US ADR debut on the NASDAQ — large enough to rank among the top five share sales of all time
Driven by demand for HBM (high-bandwidth memory) for AI; SK Hynix has reached roughly the same market cap as Samsung after coming “from far behind”
ADRs set to begin trading July 10; proceeds fund capacity expansion in South Korea and the US, plus a richer US-style PE multiple
3. $850 Billion in Future Data Center Leases
The world’s largest cloud companies have committed $850 billion in future (not-yet-built) data center leases, led by Meta and Microsoft
Microsoft added $41 billion in commitments to a total of nearly $197 billion; it earlier put leases “on ice” in 2025 and “come to regret it”
Oracle has been flat for two quarters (digesting OpenAI-related capacity) while Meta signs “as much as they seem to be able to”
4. SpaceX Record Bond Sale & A16Z Interview
SpaceX completed a record $25 billion investment-grade bond sale with $89 billion in peak orders; refinanced ~$17.5 billion of prior X/XAI debt from 9.5–12.5% down to ~5.5–6.5%
A16Z’s David George said retail took ~30% of the SpaceX IPO; firm’s stake is now valued at >$10 billion, its largest-ever return
Discussed orbital data centers (“StarMind”), framed as “airplane-sized GPU racks in space”; SpaceX equity traded higher at $158/share
5. Cerebras First Post-IPO Earnings
Cerebras stock fell ~17.5%, its biggest drop as a public company, on sequential margin contraction despite a sales beat
Record revenue of $191 million, up 92% YoY; cloud business up 167% YoY; guided full-year gross margins 10 points above plan
Margin hit of ~10–15 points came from renting back sold gear to keep up with demand; CEO Andrew Feldman noted the firm avoids HBM, CoWoS, and 3nm constraints; signed OpenAI Dec 24, in production Feb 1
6. Defense Tech & China Crackdown
Defense manufacturing startup Hadrian in talks to raise up to $1 billion, quadrupling its valuation to $7.5 billion (from $1.6 billion in January); investors include a16z, Founders Fund, Lux Capital
Alibaba sued the US Defense Department to be removed from its blacklist of firms allegedly aiding China’s military
ByteDance in early talks for a record $20 billion global loan; SoftBank’s Masayoshi Son called labeling AI a bubble “an insult”
Trends Identified
1. Vertical Integration and Compute Diversification
Frontier labs are racing to own more of their hardware stack. OpenAI’s Jalapeno chip mirrors Google’s TPU strategy and Broadcom’s belief that every major lab will eventually design custom silicon. The goal is to break dependence on Nvidia GPUs and drive down inference costs, suggesting the AI compute market will fragment into a heterogeneous mix of GPUs, hyperscaler ASICs, lab ASICs, and pioneering architectures like Cerebras’s wafer-scale approach.
2. The Memory Supercycle as Structural, Not Cyclical
SK Hynix’s massive raise, Micron’s looming earnings, and the persistent HBM shortage point to a structural shift in memory economics. Speakers argued rising manufacturing/capital intensity and the strategic value of memory mean supply-demand stays in memory makers’ favor — challenging the historical “memory is cyclical, boom-and-bust” textbook view.
3. Capital Markets as the AI Build-Out Enabler
SpaceX’s IG-rated $25 billion bond sale and SK Hynix’s $29 billion raise show that capital access — not just technology — is the defining constraint. Investment-grade ratings unlock the deep, cheap debt markets companies need to finance hundreds of billions in AI infrastructure that cash flow alone cannot fund.
4. Real Estate as the New AI Bottleneck
Both Cerebras and the $850 billion lease story underscore that data center construction — moving “at the speed of real estate” — is now the limiting factor, even as the AI market itself moves at “blistering speed.” This is reframing where value and risk sit in the build-out, and even spawning ambitions for orbital compute. —-
Sentiment Analysis
Overall Market Sentiment: Constructive but Selective
The dominant mood is bullish on AI infrastructure demand and memory durability, tempered by caution on stretched valuations, financing complexity, and margin pressure at individual names.
Risk Factors Highlighted
OpenAI chip financing opacity: The funding mechanism for tens of billions in Broadcom chips remains unclear.
Margin contraction in AI hardware: Cerebras’s 17% drop shows ramp costs can punish stocks even amid revenue beats.
Memory cycle reversal: If the historical cyclicality reasserts, current memory valuations could unwind sharply.
Data center construction bottleneck: Real-estate-speed buildout constrains compute deployment across all players.
Stretched AI infrastructure debt loads: Heavy reliance on borrowing (SpaceX, ByteDance) raises leverage exposure.
China/national-security frictions: Blacklisting, lawsuits, and IP-theft accusations cloud Chinese tech.
Hyperscaler ROI doubts: Sentiment on AI spending ROI “ebbs and flows” each quarter.
Micron print as market bellwether: A disappointing reaction could ripple across the AI trade.
This episode was covered in today’s The Market Signal — 2026-06-25, a cross-source synthesis of multiple podcast reports.