# A 35 Billion Dollar Private Credit Deal Funds Google TPUs as AWS Reportedly Sells Trainium Externally - Custom Silicon vs Nvidia - Week of June 22, 2026 > Custom Silicon vs Nvidia weekly for the week of June 22, 2026: a $35B Apollo/Blackstone private-credit deal funds Google TPUs with Broadcom backstopping the debt, AWS is reportedly shopping Trainium externally, and Jim Chanos targets the neocloud landlords. ## Custom Silicon vs Nvidia ### Week of June 22, 2026: A 35 Billion Dollar Private Credit Deal Funds Google TPUs as AWS Reportedly Sells Trainium Externally --- *Custom Silicon vs Nvidia, Weekly, Issue dated 2026-06-22* **A note on the tape:** Sixteen episodes touched our beat in the strict 7-day window (June 15–22), no 14-day fallback needed. But it was thick with *pundits* and thin on *operators*: no Broadcom, Google, Amazon, Microsoft, or Marvell silicon executive spoke, and the closest thing to an insider was a neocloud CEO who is loudly long Nvidia. Treat the design-win color accordingly. --- **TL;DR** - The one number-moving item: a **$35B Apollo/Blackstone private-credit deal to fund Google TPUs**, Broadcom backstopping the debt, Google the leases. Custom silicon now scales through the capital markets, not a benchmark. - **AWS is reportedly shopping Trainium externally**, the chip business pegged at a ~$50B run rate and Trainium 4 "sold out" a year before it ships. If true, the first credible merchant alternative to Nvidia in training. - The bear voice got smarter: **Jim Chanos** wants to be "long what the chips produce, not where the chips reside", short the neocloud "equipment-leasing companies" at single-digit ROICs. --- ## What's new Ranked by what moves a book, financing and capacity first, pundit framing last. **1. The $35B Google-TPU neocloud. [PUNDIT]** On *The Six Five* ([2026-06-15](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOj6eJtJf04FLdh2Evvnq-2FnBLAS8h-2F7guVCw5YA07mH4ZUpV2IbZGGJya-2FLSiczbouUsyKYuVtBu4QONtpLImXQIE7ZiKMf9qtdyoM1sWj-2FSSw-3D-3DIl_x_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKZr4KaWmMUYctrMoiV7W-2FutZjJ6JCHCQqCg3JMkdLk3suA109a-2BhFU1yBvFufu9TPkaNuZ4OznobznEsrh0H5Iyn-2B-2FH1VhgtXCbuhShFPwgs1WdB4MYio8XoRoSz73GsXA-3D-3D)), Patrick Moorhead (Moor Insights) and Daniel Newman (Futurum): "$35 billion Apollo and Blackstone deal… What are they going to buy? Google TPUs from Broadcom… backstopping the senior debt and Google backstopping the leases. So… this is essentially a Google TPU NeoCloud." Per Moorhead, Hock Tan framed the XPU platform as "targeting more than 20 gigawatts to compute through '28." ASIC capacity is now financed like a lendable asset class, underwriting the AVGO and GOOGL ramps. **2. AWS reportedly selling Trainium outside its own walls. [PUNDIT, secondhand]** Across same-day AI-news shows (*Founder Built*, [2026-06-19](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOipwlcN-2BEQ7g5X8c4pVrYSHbYFVGbJz5x4WS1NqNJFHVE-2FCB3MgINU55JtKPNNQY92RiZSXc6Xhv-2Bcx0ykjYa9-2F6N-2B5-2FQvtdcdsDIyPc0OVpw-3D-3Dkffy_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKTYCf2wvF-2Fm1ZijVEYrE0HymRzZV4aZ5Dz5TXZL1I9Au-2B-2FazaQbqWZI-2BwU0-2BttxVF0DmBsChqx4wQysLXRy8uhRe0T5B5dOBWTnLAinH6xhugkN9UibsvDFUd9mlSdg4cQ-3D-3D); *AI Chat*, [2026-06-19](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhbbJDhMbZyLshP97dJlzWAQc7S4XH2GNwjsYgzahmWp64jhEVnz-2FQiDNOukeNm2Yva46SQycBIuniiFH0tgpBgS7mJUETDcGpkmmjzQLg2pw-3D-3DAQ07_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKUKqwrg9pBoA0FmhJDKo6UrZjM-2FomwSKC7JikzjqNoahu-2FpCktRdlLbi60hPV2XUAeY-2FsDaAREUbF5uRdgVl3h-2FxJ18v-2FaJqNvX7p61FGQX7N32IECHq3P97HT1Z4dzq2g-3D-3D)): "AWS is in open talks… to sell Trainium chips outside of just being built inside of AWS… Andy Jassy pegged the chips business at about $50 billion annual run rate… Tranium 4 capability is already completely sold out… more than a year before this chip is actually shipping." Every one-off sale forfeits a five-year captive cloud stream, and Amazon still must pry TSMC capacity from Nvidia. I can't confirm Jassy's number from primary source, tag it reported. **3. The ASIC share number, in an analyst's framing. [PUNDIT]** On *TechSurge* ([2026-06-16](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjbY6v2MTlZXQqdrBzfED-2FIoYclvnRKVYiRg-2BLe28IlrQ3VXuEwZzoaxzf2iaf4esfYGoAoeLDzwDlDy3nkUEmdDlo4-2FRjZl8LakLsYEzPmDg-3D-3DghEE_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKTibbXY2Ucts8Ojv-2FxjS9fZVPB3FoMEOalp9lNWToSV9irgbeaA-2Fxaxl7W3-2Bhk6Nak4Kw-2BgtxR-2FWcTILxfJbqQSpSp0Wzx7-2FlZe80R7O-2BLKy-2B-2FZprglClgGRHmHwSeM4bA-3D-3D)), a semis analyst put ASICs at "mid-teens of the revenue shipments… could it be 25% or 30% of a much bigger pie? Yeah, maybe." His steel-man: ASICs win "large, stable, internally developed workloads"; GPUs keep the flexible ones. Same episode pegged Anthropic's annualized revenue at ~$9B (Dec) to ~$14B (Jan) to ~$30B (April), the demand denominator. **4. Operator reality check from a neocloud CEO. [OPERATOR]** Lambda's Stephen Balaban on *The MAD Podcast* ([2026-06-18](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjqsbrmoeAuq7DiuJR-2BrtAqU0HOURNxbwLcADZS9uufiV7bpMtXwluvu0bAy4bRUnqmv2YDowEKTucjU4nJmSrQIXYpKwqhF9tw8lLs5kRWew-3D-3Djpvg_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKbQhragv00ypIxtCzyS4Yt5Fes8DPx0o9Z5JKyG7aczJcWbBnmw0YzSzZbN54TxJYuCMwYW-2FwvaCbQ882VYhh2GurYd96TbEa0XgZ8e2OoSQ4s9EaynCY2anJwbxFr59Hw-3D-3D)): Lambda runs "everything… H100s, H200s, B200s, GB200, B300s, and VR200s coming soon," staying all-Nvidia because "they're the only chip provider… available in every single major cloud platform." His cost stack, ~$2–3B/GW power, $10–15B/GW data center, **$35–45B/GW compute**, shows where the chip fight is. And he's renting 2023-vintage H100s at a *higher* rate now than at launch. --- ## The debate **For ASICs.** The capital markets just voted, $35B of private credit for Google TPUs, Broadcom guiding to $100B AI revenue, Trainium 4 reportedly sold out a year early. Custom silicon no longer needs a benchmark win, just financing and a captive workload. **For the merchant.** The only operator buying at scale this week is all-in on Nvidia for the ecosystem ([Balaban](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjqsbrmoeAuq7DiuJR-2BrtAqU0HOURNxbwLcADZS9uufiV7bpMtXwluvu0bAy4bRUnqmv2YDowEKTucjU4nJmSrQIXYpKwqhF9tw8lLs5kRWew-3D-3Dj5aP_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKa-2BKCBzyWMwrqULnLTArDYaUdUwdKiJX880oxdqOHYu1jSUrCHqFL7f-2BzQzA5Ud-2Blumpwx8r-2FTHh8wXBUe4D78pUo-2BNclMBfScnGjgAMYxkYKX4dVVCbKv5rrcsU0XGmUA-3D-3D)). Imran Khan on *RiskReversal* ([2026-06-19](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhjVFlHrbOZTefVWSwBCUERI4JEjqR2oCTILPCk16tsyfYT5X5-2BoWjy8EanaP3mTCIaw2lfuRtcn2L-2By7fIc-2F48RwjwBoqTEB-2BdRm-2Bo53kCFw-3D-3DyXtC_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKe7NmAZuiFqO6ikjOKBcyg7Y-2BBMYBk9vPHGoBqPHoBAbRzr-2BK2GVrnCI-2FyzTIboQVKvNccBaxS1TZlCWiZGlGfeiv2yI6WXrDrKCoxxOJoTFnL7zpd18WT9FSI5ZCAWR6Q-3D-3D)) argues Nvidia at ~16x next-year earnings is "not the Cisco multiple we saw in 1999", the market already prices 80% growth halving, while CUDA stays the moat. His one worry: "their biggest customers are also their biggest competitors." **The third voice.** Jim Chanos on *Monetary Matters* ([2026-06-20](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOg-2BZ-2FbiAlCcdQYF0bsFVQ6pgfy2VpfJA-2BSnjfodH-2F-2BDSp-2BBLb9ZGFEj410hF6RueqKg66yeYzTUIwtVwXa4b6NJsGWxSS-2BUR1uN8PDkj4bzuw-3D-3Dp0Xf_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKZf8uCNuqIonKgZXLYHcUjYROWjOjQNUBLhnN8lBdbzdtsjNZcYeBw8AQE3mQw7XEhgoNflz5xclZfHf7Ish9AJWiw7rO38SroYRIC75NuJ-2BUlSQg1Y5Qqwv0X-2FYVqx8NA-3D-3D)) sidesteps GPU-vs-ASIC: neoclouds are "equipment-leasing companies… finance companies in effect," penciling out at "single-digit ROICs pre-tax" even on a generous 10-year GPU life, a 1998–2001 echo where picks-and-shovels book revenue while hyperscalers capitalize the spend. Co-panelist Val Zlatev's live datapoint: GPU rental prices "down 20–30% year on year… into December" are now "up 40, 50%, even more." **Where I come out this week:** the ASIC bull case got more real via *financing*, not performance, a quieter, more durable form of share gain than any MLPerf headline. But the operator writing the checks still buys Nvidia, and Chanos is right about who holds the depreciation when the rental spike normalizes. Long the silicon (NVDA, AVGO), skeptical of the landlords. --- ## Stocks in play - **NVDA**: *Bull:* the only chip in every cloud; ~16x forward on 80% growth ([Imran Khan](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhjVFlHrbOZTefVWSwBCUERI4JEjqR2oCTILPCk16tsyfYT5X5-2BoWjy8EanaP3mTCIaw2lfuRtcn2L-2By7fIc-2F48RwjwBoqTEB-2BdRm-2Bo53kCFw-3D-3DT1Qm_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKRzQzDJ76tjYVsDFiQdgHNFkNTK21mQ-2Fu9m2GmAfbFNjEny7y0-2BVBKYB6CwwYNDbxtnKxxV0j-2FjZyfX6MFJtMAeURmYnzslYnHpGGZ5AhlOGZL3JOgAMT-2BOgt4ZZuWylrQ-3D-3D)). *Bear:* customers are competitors; $20B debt raise despite ~$100B FCF. *Watch:* any MLPerf submission from the ASIC camp (still absent). - **AVGO**: *Bull:* backstopping the $35B TPU financing; >20GW XPU platform; $100B AI revenue guide. *Bear:* no fresh customer-list disclosure. *Watch:* OpenAI/Apple XPU confirmation. - **GOOGL**: *Bull:* TPUs now a financeable, multi-gigawatt asset class. *Bear:* needs $80–85B fresh equity to fund capex ([David Woo](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOgXa9gT1kOJP4z9zFbPneLN1sxQJ6QpqxdNS3h-2F-2BTSP-2FYzSxpdG1SlhnLX6fkZVMePfraeZ-2FBeHNy0FY6Qjf-2BrsvQDfbUuGqiPR7OTpbHMx9Q-3D-3Do41G_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKTlduf4DujqI1r-2FTmp5u9kpujDdkCWT70-2Fp2lgTF8pmhmi4CiiKvIcqLjk3asBFUci17IXl9zuAxs76HkKuzjUzkTvVdX8CzqUFCWLXTFMkIvmKGsKwUsF2V3-2F2ZtPGF3A-3D-3D)). *Watch:* TPU v7/Ironwood customers beyond Anthropic. - **AMZN**: *Bull:* Trainium reportedly ~$50B run rate, sold out a year early. *Bear:* secondhand, and external sales cannibalize captive cloud. *Watch:* primary confirmation. - **MSFT / META**: *Bull:* Meta's walled-garden data advantage compounds "as chips get commoditized" ([Investor's Podcast](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhkHhwwH3e01spua-2FQXqOYIvjPr6U3p8CU7q9MlF5TgMdppLsSrtcq8kh4ozNNdPhpeOxyaTy1x6Q3AyVa8d-2F7Sn2V-2BLn6QiaemsEHXbK4Fmw-3D-3DTJbi_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKRYmstcHZubB-2BkVlPlj-2FxOmp3xsJfaCsEpkPGxuWfWiNJgz5lQ7Da-2FoQOQtNgzPmedDleKcBCVqZCC9PiHFj07tkVjTLI0g6mk9-2BvyLrVWK6XdZXR4aqTWmB3IM-2BYbnSZg-3D-3D)). *Bear:* zero Maia/MTIA volume color; Meta not cheap on FCF. *Watch:* a Maia 2 / MTIA v2 metric. - **ARM / ALAB**: *Bull:* Neoverse and connectivity ride inside every XPU rack. *Bear/Watch:* no new royalty-mix or design-win color, watch attach rates as the $35B TPU capacity lands. - **INTC**: *Bull:* an AI-generated digest reported Google committing ~3M TPUs to Intel's fab for 2028 plus a BofA sell-to-buy double-upgrade ([Telltales](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhqIC2s3swzB-2F9oz8zGHSi4GzfEh8RAu-2BZh2wRt03S1syyiKXqdNtJZNvVu9-2BkVxicZqbYEe-2FswD5aYlVnPNQVKKD96Q-2F1wmwHSWfujiHkOTA-3D-3D8iwg_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKYbGxkX8TICWvCw-2FOfaaHP2atJdOCCF1vHjE74-2BehL9yKuk1EnvgquqDj1kgc1VO7KJOvM4JMCuk8WDqlBq-2Bhu7T6enA2cUB4t-2FqwiWOt42CgMaTcZiX2ovKgWERQJgYTg-3D-3D)). *Bear:* Newman calls it "a complete nothing burger", Intel gets *packaging/IO-die*, not the wafer ([Six Five](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOj6eJtJf04FLdh2Evvnq-2FnBLAS8h-2F7guVCw5YA07mH4ZUpV2IbZGGJya-2FLSiczbouUsyKYuVtBu4QONtpLImXQIE7ZiKMf9qtdyoM1sWj-2FSSw-3D-3D9GNo_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKalPa-2BrLDr-2BhuYM3rVpjfMKiXmEkRvd6fsD5bs5jvy1hIYkyu6qHPgPqn8rrdmUa9KIqdNiX4ACmiJhFf-2BWy0KTRyGimS-2FeGNFkxusB6dwwCMX12lkgaFkuVYIOA8OfJlQ-3D-3D)). --- ## Read-throughs - **HBM, Micron / SK Hynix / Samsung.** Micron reportedly certified for Nvidia's HBM4, FY production spoken for ([Telltales](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhqIC2s3swzB-2F9oz8zGHSi4GzfEh8RAu-2BZh2wRt03S1syyiKXqdNtJZNvVu9-2BkVxicZqbYEe-2FswD5aYlVnPNQVKKD96Q-2F1wmwHSWfujiHkOTA-3D-3DVYBu_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKX-2BtN-2FQ6gaLZC8m9bkuWt-2B29qxzuyHXlvAMGj-2F4o9FlhmRY-2BEec-2B2dZbBSESByA8xx2bce8ZrGqVppNiijv3VZdcSR-2BYyda7OwSI41jrxQatm-2FIk1Nap54BPUVmVBbR-2BAA-3D-3D)). Counterweight: David Woo calls the spike "an income transfer from Microsoft to Micron" and the pricing power "short-lived" given Chinese DRAM entrants ([Monetary Matters](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOgXa9gT1kOJP4z9zFbPneLN1sxQJ6QpqxdNS3h-2F-2BTSP-2FYzSxpdG1SlhnLX6fkZVMePfraeZ-2FBeHNy0FY6Qjf-2BrsvQDfbUuGqiPR7OTpbHMx9Q-3D-3DkpLw_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKdL5v6J48zKNh1771Clyl5VbusZUcRgdRB6CAK-2FtvhWMNmly7ksTkh6IjuNCi1PqMDqbUyN1z-2FGQN6xJ4SKC-2FX1dLiLHC80cdmIMBOUabJnhcTNBXuv5-2F3otJd15chInoA-3D-3D)). - **TSMC.** Still the binding constraint, Amazon must pry CoWoS capacity from Nvidia, now "Apple's replacement as the foundry's largest customer" ([Founder Built](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOipwlcN-2BEQ7g5X8c4pVrYSHbYFVGbJz5x4WS1NqNJFHVE-2FCB3MgINU55JtKPNNQY92RiZSXc6Xhv-2Bcx0ykjYa9-2F6N-2B5-2FQvtdcdsDIyPc0OVpw-3D-3DkwmJ_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKczPobx58Eg-2FMBeXjkSD1P9wP3jvGzuBGsonMCKsdAcyYHot40DMBaG3hRmopZVjk-2F8wtjGh9HkZ82Kt0h-2BbMl44EpP3X4ZctzHreBRVoXJqFo2YgIChE5tcxg-2BoG4qSoA-3D-3D)). - **Connectivity (ALAB / MRVL / AVGO).** Networking is "as important as the compute" ([TechSurge](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOjbY6v2MTlZXQqdrBzfED-2FIoYclvnRKVYiRg-2BLe28IlrQ3VXuEwZzoaxzf2iaf4esfYGoAoeLDzwDlDy3nkUEmdDlo4-2FRjZl8LakLsYEzPmDg-3D-3DOCCr_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKUWUNi-2BpAbj-2FKbHhlA5-2Bj41znqgIQ5sX2hdGIWyhS5GRE2t3y14gSxYA72jQfdhNCXxjw8kF2nrFvDXskIQh85qCL5uUKZxr-2FbzssTyscp-2FA5XF3Fcv7Sy20-2B0P6D4Sk8Q-3D-3D)); NVLink in-rack, InfiniBand/Ethernet across racks. No fresh Marvell disclosure. - **Alt-silicon, Cerebras (CBRS).** Now public at ~$46B, with reported ~$20B of OpenAI orders and a wafer-scale pivot to inference ([7investing](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOhSZA2MKMDNmjt-2BmFyIZWeXmWQQKfJ2gFG6zJjkrTA-2F49q0BojXT1d9FAS15oMSxiLCjiS9AR76v3pdeMvXsl7ZlXTXkKsj4eX9Ke7XBIjz1A-3D-3DFl-7_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKUvFuoRz0fmucxIpx7vekwBbAyRI70zf9q5KdOHe9sKrhTuF8Sv-2Bz7kQU4TFoRsHjMXPH8n6roA18z8DMtLTcZ0jbofRqKUAmzOzU7gQL-2BkWTlLW67GArov3xkNu3cUVBg-3D-3D)). Note the warrant-driven circularity. - **Capex backdrop.** Hyperscaler capex is now ~135% of operating income for the top five ([David Woo](http://url7324.matterfact.com/ls/click?upn=u001.idHmPrr2Geh7KYLAsTy7NkrIVb-2FgA4pmf2rMXQwGcOgXa9gT1kOJP4z9zFbPneLN1sxQJ6QpqxdNS3h-2F-2BTSP-2FYzSxpdG1SlhnLX6fkZVMePfraeZ-2FBeHNy0FY6Qjf-2BrsvQDfbUuGqiPR7OTpbHMx9Q-3D-3DHfcV_7mLGwmUci-2BLaXswv9WX1yTgqn3Wad-2FotHhzHgSNAZbXq9Mlyqd8YjgBozVwG-2FKaYbO5WTqa3HBuEaRHpGY5PKSEyjYt4UXmxeUMZF62qhLSkU-2Ft9Gqdsgs0f-2FpT-2B-2BwgCp2voiRSNuk7Af5bIPs7zwy87VbGD9HsMnMvk6BPsKQ9LTmUh8XGgKqa-2FRBCf32EcZpgoytCumVSAzToRJXkmtw-3D-3D)), the deep-pockets era is over, which is why the financing matters. --- ## What changed vs last week Last week (June 1) the ASIC bull case rested on a *performance/share* claim, Feldman saying CUDA had lost 70% of frontier training. This week it moved from the benchmark to the **balance sheet**: the new evidence for ASIC scaling is a $35B private-credit structure for Google TPUs. Three deltas: a heavyweight *short* voice entered (Chanos on the neocloud middlemen, absent last week); operator color flipped pro-Nvidia (Balaban) vs last week's ASIC-leaning Feldman; and GPU rental prices now have a number, up 40–50% since January. Still missing, as last week: any MLPerf benchmark from the ASIC camp, and hard MTIA/Maia/Marvell/Alchip disclosure.