Mad Hedge Biotech and Healthcare Letter
April 16, 2026
Fiat Lux
Featured Trade:
(THIS BIOTECH MIGHT HAVE CRACKED THE UNDRUGGABLE CODE)
(RVMD), (MRK)
(CRWV), (META)
Last week, a single Meta (META) employee consumed 328.5 billion tokens in 30 days using Claude Opus 4.6. At published rates, that tab runs $1.9 million. One employee. One month. And Meta didn't flinch.
If you want to understand why CoreWeave (CRWV) just signed two of the largest compute contracts in the short history of AI infrastructure, start there.
The AI bubble crowd has been wrong for three years running, and they're about to be wrong again.
The bear case on CoreWeave always rested on one assumption: that demand for raw compute would plateau as AI companies shifted toward leaner, cheaper inference models. DeepSeek scared everyone in January. The hand-wringing hasn't stopped since.
What the skeptics missed is that Anthropic's new Mythos model, not yet publicly available, is reportedly 5 to 10 times larger than any model ever trained. Pre-training scale, the thing everyone declared dead, is very much alive.
And models of that size don't run on wishful thinking; they run on GPU clusters the size of small cities.
Which is exactly where CoreWeave comes in. Last week, the company expanded its infrastructure agreement with Meta through December 2032, a capacity deal worth $21 billion.
In the same week, CoreWeave inked a separate multi-year arrangement with Anthropic to bring additional compute online starting later this year.
Two of the most aggressive spenders in AI, locking in capacity through the end of the decade, in the same seven-day window.
The financing picture, which spooked investors through most of 2025, is improving faster than the market has priced in.
CoreWeave carries 6x adjusted EBITDA leverage based on FY2025 figures, a number that looks alarming in isolation. Context matters, though.
The company primarily uses asset-backed financing to purchase GPUs rather than the convertible notes that have blown up so many high-growth names before it.
The take-or-pay contract structure means customer prepayments offset a meaningful portion of the upfront capital outlay.
Credit markets are already voting with their feet; spreads on CoreWeave's five-year credit default swaps fell 23% in a single week after the company unveiled an $8.5 billion investment-grade loan facility. When the bond market relaxes, pay attention.
The real insight here sits inside the token economics. Anthropic's revenue has run from $9 billion annualized at the end of 2025 to $30 billion today, even though Claude Opus 4.6 costs roughly 2.6 times more per token than OpenAI's GPT-5 equivalent.
Customers are paying the premium without negotiating. That tells you the bottleneck is capability, not cost.
When AI companies stop haggling over price, the infrastructure providers supplying the underlying capacity operate in a seller's market, and CoreWeave is one of the few scaled sellers that exists.
Still, the risks remain and shouldn’t be glossed over. Execution delays at this leverage ratio are not a nuisance; they are an existential problem.
Vertical integration across the full stack, the only sustainable moat against the hyperscalers long-term, remains unfinished business.
And if the financing market seizes up for any reason, the equity cushion compresses fast. Position size accordingly.
That said, CRWV has recovered most of its early 2026 losses and is pressing against the $115 resistance level that capped every rally since the IPO. The consolidation above $65 has held.
The demand signals coming in from both Meta and Anthropic suggest the market is now baking in a structurally higher CapEx regime through 2027 and beyond.
A decisive break over $115 opens a clean run toward $150, and the January 2027 $120-$130 call spread, currently trading around $4.50, offers a reasonable vehicle for the patient money.
The AI buildout is nowhere near its peak. The customers are not flinching. The only question is whether you want to be the one holding the bill or the one sending it.
One Meta employee spent $1.9 million on tokens last month, and nobody at headquarters lost any sleep. Imagine what the other 70,000 employees are running.
Global Market Comments
April 15, 2026
Fiat Lux
Featured Trade:
(THE GOVERNMENT’S WAR ON MONEY)
(TESTIMONIAL)
"Sometimes we stare so long at a door that is closing that we see too late the one that is open," said Alexander Graham Bell, inventor of the telephone.
Global Market Comments
April 14, 2026
Fiat Lux
Featured Trade:
(THE IRS LETTER YOU SHOULD DREAD),
(TESTIMONIAL)
Just a quick note of appreciation for your helping me decide to get my clients into (GLD) with a 15% allocation early this year.
It sure has helped me to be more of a hero to my clients this year than a goat...
Best wishes to you and yours! Keep 'em coming.
Brad
Bakersfield, CA.
Global Market Comments
April 13, 2026
Fiat Lux
Featured Trade:
(MARKET OUTLOOK FOR THE WEEK AHEAD, or WHAT TO BUY AT THE BOTTOM),
(AAPL), (GOOGL), (AMZN), (NFLX), (PANW), (ZM), (GS), (MS), (JPM), (BAC), (C), (BA), (CAT), (FCX), (WMT), (COST),
(DAL), (RTX), (FXA), (SLB), (CCJ), (GLD), (GDX), (NEM), (WPM), (XOM), (OXY), (NVDA), (AMD), (MU)
"The next leg of the bull market is going to have to be a fundamental leg, not a liquidity leg, and investors are in a little bit of a malaise trying to figure that out," said Anthony Scaramucci of Skybridge Capital.
A Crypto Upstart Claims to Beat GPT — and the Market Believed Every Word
OpenServ's SERV token surged 126% after its Nano LLM reportedly outperformed GPT-5.4 on cost and speed. The benchmark claims are unverified. The price move was not.
April 10, 2026 | Mad Hedge Fund Trader
In AI infrastructure investing, benchmark claims have become their own form of currency. Announce that your model beat GPT, watch the token spike, and dare the skeptics to prove otherwise. That is exactly what happened last week when OpenServ, a crypto-native AI agent platform, dropped performance data suggesting its SERV Nano language model matched or outperformed OpenAI's GPT-5.4 at 20 times lower cost and three times the speed.
The result: the SERV token surged 126% in a single week. The AI infrastructure trade is alive, it is loud, and it is moving fast.
What OpenServ Is Building
OpenServ operates an AI agent platform built on blockchain rails. The core thesis is straightforward: autonomous AI agents need financial infrastructure — payment systems, identity protocols, coordination layers — that traditional finance was never designed to provide. Crypto-native infrastructure can fill that gap.
The SERV Nano announcement centered on the company's BRAID framework — Bounded Reasoning for Autonomous Inference and Decisions. The claim: SERV Nano delivers comparable reasoning performance to GPT-5.4 at a fraction of the cost and at higher throughput. The specific numbers put forward were a 20x cost advantage and 3x speed improvement.
If those numbers hold up under independent scrutiny, they matter. GPT-5.4 sits at roughly $2.50 per million input tokens and $10 per million output tokens. A system delivering comparable results at 20x lower cost would meaningfully expand the economics of agentic AI deployment at scale. That is real money for anyone running autonomous agents across financial workflows.
Read the Fine Print
Here is where experienced investors need to pause. Benchmark comparisons in AI are not always what they appear. When a third-party framework claims to match a frontier model, the performance lift can come from several sources that have nothing to do with raw model superiority.
Task selection is one. A narrowly framed benchmark that plays to a smaller model's strengths tells you very little about general capability. Smart routing logic is another — a system that sends easy queries to cheap models and hard queries elsewhere can lower average cost dramatically without the underlying model being competitive. Deterministic scaffolding can reduce output variance and inflate benchmark scores without improving actual reasoning depth.
As of this writing, independent third-party replication of OpenServ's results has not been confirmed. OpenAI's own documentation already positions GPT-5.4 nano as a budget-tier, high-volume option — which raises the question of whether this comparison is frontier versus frontier or budget versus budget with better marketing.
None of this means the claims are wrong. It means they require verification before you act on them.
Why the Market Moved Anyway
The 126% surge in SERV is not primarily a story about a verified technical breakthrough. It is a story about how crypto markets process AI narratives in 2026.
Institutional interest in decentralized AI infrastructure was already building before OpenServ's announcement. The same week, Grayscale filed an amended S-1 for a Bittensor Trust — a significant signal that institutional capital is beginning to engage with decentralized AI networks in a structured way. DeXe, another AI infrastructure token, gained 14% in the same period. The sector was already in motion.
What the SERV announcement provided was narrative clarity. A specific, quotable technical claim — beats GPT at 20x lower cost — gave momentum traders something to attach to the price move. In crypto, the story often moves first. The proof follows later, or it does not. Experienced traders know the difference.
The Cost War Is Real Even If the Claim Is Not Proven
Here is what I know for certain: the underlying competitive pressure driving OpenServ's claims is entirely real, regardless of how their specific benchmarks ultimately hold up.
AI inference costs have collapsed. DeepSeek V3.2 delivers roughly 90% of GPT-5.4's performance at approximately one-fiftieth the price. Anthropic cut Opus pricing by 67% in early 2026 to stay competitive. Google's Gemini 2.5 Flash Lite enters the market at $0.10 per million input tokens. The floor has dropped so fast that a model priced at $0.50 per million tokens now looks expensive by comparison.
This matters for the agentic AI buildout in finance. Autonomous agents running fraud detection, compliance monitoring, deal sourcing, and portfolio rebalancing need to process enormous volumes of tokens continuously. Every order of magnitude reduction in inference cost opens up use cases that were previously uneconomical. That is the genuine market opportunity — and it is large.
How to Think About This Trade
The AI agent infrastructure space will produce real winners over the next several years. The companies that build reliable, auditable, cost-competitive agent infrastructure on whatever rails — centralized or decentralized — will capture durable value. The projects that ride the narrative without the technical substance will not.
For SERV specifically, the questions I want answered before treating this as a fundamental play rather than a momentum trade: Has the BRAID framework been independently benchmarked? What production deployments exist? Is the token capturing value from real infrastructure usage or primarily from narrative-driven demand?
For the broader AI infrastructure trade, the story is more straightforward. The cost compression is real. The demand for agentic AI in financial services is real. NVIDIA's sixth annual State of AI in Financial Services survey found that 65% of institutions are now actively deploying AI, up from 45% just one year ago, and nearly every respondent plans to maintain or increase AI budgets in 2026. That is not a narrative. That is a structural shift in how the financial industry operates.
The Bottom Line
OpenServ's claim that SERV Nano outperforms GPT-5.4 at 20x lower cost drove a 126% token surge in one week. The claims are unverified. The underlying trend — relentless compression of AI inference costs by challengers from every direction — is rock solid. The AI infrastructure buildout in finance is not slowing down, and the market knows it.
Separate the verified trend from the unverified claim. That is where the real trade lives.
This is an excerpt. You do not have sufficient rights to view the full content.