#160 KI-Blase, Nebius, Cerebras, ChatGPT, Sehr sehr gute Nachrichten

Summary

Theses (episode spine)

• Broad market rally continues: S&P 500 and Russell 2000 both hit new all-time highs last week, and Russell 2000 strength signals the rally is broadening beyond mega-cap tech into small/mid-caps, the “backbone” of the US economy (Max).
• July inflation data came in cool (CPI +0.1% MoM, headline 3.4% YoY, core 2.5%), pushing the market-implied odds of a September rate hike down to 32% from 52% before the data (Max).
• Nvidia announced a roughly $500B compute-financing partnership with major asset managers (including Blackstone and Apollo) to mobilize third-party capital for AI infrastructure, using chips and data centers as loan collateral (Max).
• Max cites a Zero Hedge article comparing today’s AI buildout to both the 2000 dotcom bust (equity-funded, patient capital) and the 2008 subprime crisis (debt-funded, negative convexity), arguing AI infrastructure financing now resembles the 2008 credit-leverage structure more than dotcom; the article’s key warning is that a mere deceleration in growth acceleration (the “second derivative”), not an outright demand collapse, could be enough to strain the leveraged financing structures.
• Max says the article gives him “another lens” to watch the numbers through but is not yet changing his portfolio because the deceleration the article warns about is not showing up in current results; Stefan is more unsettled, specifically flagging Jensen Huang’s proposal to use chips as loan collateral as a red flag, comparing it to Enron’s securitization of broadband capacity, since chips (unlike real estate) lose most of their value within a few years.
• Nebius Q2 beat across the board: revenue $582.3M (guided $557-573M, +454% YoY), adjusted EBITDA $236.2M (guided $157.9-175M), ARR $3B (+58% QoQ), and 2026 contracted capacity raised to 5GW from 4GW; Max calls it validation that Nebius, and the broader neocloud/AI-cloud sector, is “on the right track.”
• Both hosts rate Nebius a hold, not a buy, after its ~68% three-week rally off the late-July Aschenbrenner-fund-liquidation lows (from ~$148 to ~$240-248); Max would add again around $200 or lower, and Stefan likewise flags ~$200 as his target re-entry level, warning listeners to size AI-infrastructure positions small given the stock’s history of 40-50% multi-week drawdowns.
• Cerebras’ core (non-GAAP) Q2 results beat across metrics (core revenue $209-221M vs. $194M consensus, core EPS -$0.05 vs. -$0.18 expected) despite a GAAP revenue miss driven by one-off items (pass-through data-center costs, warrant amortization, IPO costs); management raised FY2026 core revenue guidance to $880-890M from $855-865M and guided for 2027 revenue to more than triple.
• Max bought the post-earnings ~16-17% Cerebras selloff at $219, framing the stock as a long-term inference/cloud play with structural supply-chain advantages (no dependence on HBM, CoWoS packaging, or 3nm nodes) and new large customers OpenAI (GPT-5.6 inference) and AWS (Trainium-based inference from Q1 2027); he would add more around $170-180.
• Stefan does not own Cerebras — he already carries roughly 40% of his portfolio in AMD and avoids adding more semiconductor/AI-infrastructure exposure — and says he is tactically bearish on the AI-hardware/chip sector for the next ~1.5 months, expecting further corrections in names like Nebius and Cerebras.
• The Iran/Strait of Hormuz standoff has become an economic stalemate rather than active conflict; Brent crude sits near $91-92, keeping pressure on bond yields and inflation, and Stefan argues the US has largely exhausted its effective military and economic leverage (finite missile stockpiles, damaged regional bases), expecting the standoff to persist at least until the November midterms.

$NBIS (Nebius Group)

Convergence / divergence: Max and Stefan agree Nebius delivered a strong beat-and-raise quarter and both remain long-term bullish, but both are disciplined on price — neither is chasing the post-earnings rally and both independently cite ~$200 as their preferred re-entry level.

Speaker calls:

• Max (bullish, long-term): Nebius is now his largest portfolio position, up roughly 370-390%; Q2 beat across revenue, EBITDA and contracted 2026 capacity (raised to 5GW from 4GW); he is not chasing the current rally but would add again around $200 or lower.
• Stefan (bullish, long-term): Remains long-term positive on the growth and acceleration he sees in the numbers, but rates the stock a clear hold, not a buy, at current levels; would look to re-enter around $200, and warns the stock has a history of 40-50% multi-week drawdowns so position sizing should stay small.

Cross-check:

• Price: $218.46 (Aug 19, 2026). P/E ~847 (GAAP, near-breakeven quarterly EPS distorts the ratio — company is still unprofitable on a net-income basis). Market cap: $32.5B. Next earnings: 2026-11-10.
• Recent headlines worth knowing: Q2 2026 beat on EPS (-$0.12 vs -$0.82 est.); stock fell >5% pre-market after announcing a proposed $4.5B convertible senior notes offering to fund infrastructure buildout.
• ⚠️ Inconsistencies: Current price ($218.46) is below the hosts’ stated at-recording price ($240-248), consistent with the pullback they discussed and closer to their $200 target re-entry zone. The new $4.5B convertible note issuance (dilution/leverage) was not mentioned in the episode.

$CBRS (Cerebras Systems)

Convergence / divergence: The hosts diverge here — Max actively bought the post-earnings selloff and is building a long-term inference/cloud thesis, while Stefan stays on the sidelines purely for portfolio-concentration reasons (heavy existing AMD exposure) rather than a fundamental disagreement about Cerebras itself; both agree the stock is highly volatile and that the AI-chip sector broadly could see further corrections.

Speaker calls:

• Max (bullish, long-term): Bought the post-earnings 16-17% selloff at $219 after core-business results beat consensus (revenue $209-221M vs $194M consensus, core EPS -$0.05 vs -$0.18 expected) and FY2026 guidance was raised to $880-890M; sees Cerebras shifting toward an inference/cloud-driven business with structural supply-chain advantages (no HBM, CoWoS, or 3nm dependency) and new large customers OpenAI and AWS; would add more around $170-180.
• Stefan (neutral, no position): Does not own Cerebras since he already carries roughly 40% of his portfolio in AMD and avoids adding more AI-chip exposure; is tactically bearish on AI infrastructure stocks generally for the next ~1.5 months and expects further corrections in names like Cerebras and Nebius.

Cross-check:

• Price: $197-210 (fluctuating Aug 21-22, 2026). P/E: N/A (unprofitable). Market cap: ~$49.6-50.5B. Next earnings: not confirmed; Q2 2026 was reported Aug 12, 2026 (revenue $210M vs $193.6M consensus, beat), next report likely Q3 2026 around November.
• Recent headlines worth knowing: Q2 2026 revenue beat consensus ($210M vs $193.6M); stock has been highly volatile since the print, swinging between roughly $195 and $220 over the following days.
• ⚠️ Inconsistencies: Current price range ($197-210) is close to Max’s stated buy price ($219) and recording-time price (~$220), consistent with what the hosts described. Market cap ($49.6-50.5B) is notably higher than what the episode’s revenue figures alone would suggest for a company with ~$210M quarterly core revenue, underscoring the high growth-multiple valuation risk both hosts flagged.

Topics discussed

Is the AI buildout a bubble? (Zero Hedge CDO/CCO comparison)

Summary: Max walks through a Zero Hedge article comparing today’s AI infrastructure buildout to the 2000 dotcom bust and the 2008 subprime crisis, arguing it structurally resembles 2008 more (debt-funded, take-or-pay contracts, negative convexity) since Nvidia’s new ~$500B compute-financing plan uses chips and data centers as loan collateral, similar to mortgage-backed securitization. The article’s central warning is about the “second derivative”: a mere deceleration in the rate of growth acceleration, not a demand collapse, could be enough to strain the leveraged financing structures, similar to how house-price acceleration (not the price level) turned down before the 2008 crash. Hyperscaler AI capex growth is shown slowing from 81% (2025) to 77% (2026, first deceleration) to a projected 52% (2027).

Speaker views:

• Max: Treats the article as “another lens” for watching future quarters — he would reduce positions if he sees revenue/earnings acceleration decelerate faster than expected even while headlines stay positive — but says the deceleration described isn’t visible in the numbers yet, so he is not changing his portfolio today.
• Stefan: More skeptical: specifically flags Jensen Huang’s proposal to collateralize chips (which depreciate within a few years, unlike real estate) as a red flag, comparing it to Enron’s securitization of broadband capacity, and says banks like Goldman and BlackRock are willing to package and sell this risk because it isn’t their downside if it goes bad. Notes Zero Hedge is a long-time “permabear” source but says that doesn’t make every article wrong.

Potential impact: If hyperscaler capex growth continues decelerating as projected for 2027 (52% vs. 77% in 2026), the hosts warn this could pressure highly leveraged AI-infrastructure financing structures and trigger a broader “waterfall” correction across AI-adjacent equities, similar to how credit stress preceded the 2008 crash.

Iran / Strait of Hormuz economic standoff

Summary: The US-Iran conflict has shifted from military exchanges to an economic standoff over shipping through the Strait of Hormuz, with only one or two vessels passing per day. Iran is negotiating with Oman over control of the strait and is holding out for sanctions relief and an end to the US blockade before returning to talks. Trump has threatened Oman militarily if it cooperates with Iran. The US is reportedly expanding oil investment in Venezuela (already taking over 50% of its oil exports) to diversify away from Hormuz-dependent supply.

Speaker views:

• Stefan: Argues the US has effectively run out of military and economic leverage — missile/cruise-missile stockpiles have finite shelf lives and can’t be indefinitely replenished, Iran has destroyed or damaged several US regional bases, and none of the US’s stated eight war goals (securing nuclear stockpiles, regime change, reopening the strait) has been achieved despite six months of conflict. He expects the standoff to persist at least until the November midterms, when falling approval ratings could pressure the administration to de-escalate, and calls the newly announced “Operation Economic Fury” sanctions unlikely to work given 47 years of prior sanctions.
• Max: Says he remains tactically cautious (“a bit bearish”) on the broader market impact of the standoff, since the strong rally since the late-July Aschenbrenner-fund liquidation may need “more fuel” to continue and could see a healthy 10-15% correction; flagged Brent crude at $91-92 as a factor that, if it stays elevated, pressures both inflation and sovereign bond yields.

Potential impact: Elevated oil prices (Brent ~$91-92) driven by the closed strait are cited as a factor keeping bond yields and inflation expectations higher, with knock-on effects for consumers, corporate financing costs, and government borrowing costs broadly.

Ifo Institute VAT reform proposal (Schmutz der Woche)

Summary: Ifo Institute president Clemens Fuest proposed abolishing Germany’s reduced 7% VAT rate (mainly applied to food) in favor of a single unified rate, offset by direct credits/rebates for lower-income households (estimated at 7.2 billion euros to compensate the lower-earning half of the population). Both hosts strongly criticize the proposal.

Speaker views:

• Stefan: Calls the proposal an attack on the middle class, arguing that a gross income of 55,000 euros/year (roughly 3,000 euros net/month in a major city) is being mischaracterized as “high income” when it barely covers rent, transport, and childcare; he would instead lower VAT on all goods to 7% rather than raising food prices, and criticizes the added bureaucracy the rebate scheme would require.
• Max: Says the proposal surprised him given Fuest’s previously well-regarded commentary, and agrees it targets the wrong group (the middle class) rather than achieving its stated goal of easing the state budget.

ChatGPT removes free-tier usage limits (Luna model)

Summary: OpenAI’s new ChatGPT 5.6 update introduces two tiers: “Sol,” a paid model for harder tasks (which runs on Cerebras infrastructure), and “Luna,” a free model with no usage limit, unlike competitors such as Gemini which reset usage limits every 8 hours. Stefan tested the free Luna model against his paid Gemini Pro subscription on tasks like summarizing weekly crypto news (Solana, Ethereum) and found ChatGPT’s free tier gave more complete answers in several tests, prompting some of his subscribers to consider cancelling paid subscriptions.

Speaker views:

• Stefan: Recommends listeners try the free, unlimited ChatGPT Luna model, saying it outperformed his paid Gemini Pro subscription in several head-to-head tests and could be a useful, cost-free tool for people less familiar with AI tools (e.g., older relatives).