My Thoughts on the Economy

There has been plenty of news in healthcare, industrials, and technology. Today, I’ll spend some time reviewing each of these sectors and trends I’m observing across public and private companies in them.

Healthcare This includes pharma/biotech, medical devices, healthcare services, health insurance, health IT, and life sciences tools.

GLP-1s reduce appetite and help people lose substantial amounts of weight. They were originally developed for diabetes, but they are now widely used for weight loss. Some examples are Ozempic, Wegovy, Mounjaro, and Zepbound. Novo Nordisk and Eli Lilly are two dominant manufacturers who both have been concentrating a large portion of the pharmaceutical industry growth in obesity and metabolic disease treatment. There are two things to note. One, GLP-1 pills are coming – oral ingestion is lower friction than the standard injections now. Two, the government is applying pressure on pricing. Medicare/Medicaid are trying to secure lower prices, which would improve access but naturally reduce profit margins. If people lose significant weight through medication, then I imagine bariatric surgeries will trend down, as may orthopedic devices sold by companies like Stryker and Zimmer Biomet (e.g., knee and hip replacements). If the average overweight-prone person becomes less heavy, then they would have less stress on their joints, thereby avoiding or delaying orthopedic devices or procedures.

In fact, health insurers are paying for expensive GLP-1 drugs now because they hope to save money later. If they pay for costly medications, then they would have to pay for fewer cases of Type 2 Diabetes, heart disease, kidney disease, and other obesity-related conditions. This could drastically reduce overall healthcare spending. I’d consider playing a long macro theme on wraparound services companies since healthcare plans are increasingly pairing GLP-1 medication with additional support in the form of nutritional counseling, dietitian access, etc. It will be very interesting to track GLP-1 usage once pills become mainstream; right now, about 1 in 8 Americans already use them.

AI is becoming a core infrastructure technology in healthcare – kind of how EMRs (electronic medical records) became essential in the last decade. On the clinical side, we are already seeing less physician paperwork, better workflow efficiency, and more consistent patient follow-up from private companies like Epic, Nuance, and Abridge. On the research side, drug discovery is getting accelerated. I’m particularly interested in Chai Discovery, which recently raised its Series B at a multi-billion dollar valuation and is having its AI model used by Pfizer and Eli Lilly. On the business side, we can anticipate lower administrative costs, smaller back-office staffing needs, and overall higher productivity for hospitals and healthcare organizations.

Payer and reimbursement pressure is another force to deal with. Medicaid, of course, is the government insurance program covering many low-income Americans. Trump’s One Big Beautiful Bill Act (signed in on last year’s independence day) reduces federal Medicaid spending substantially over the next ten years – there are more work requirements for adults, more frequent eligibility checks, and narrower eligibility for certain immigrant groups. The practical effect is that some people who receive coverage may lose it in time. This really matters for hospitals. In general, commercial pays the most, medicare less and medicaid even less of hospital bills. Uninsured patients pay little or nothing. More people losing Medicaid coverage means hospitals will face more uncompensated care, more bad debt, and less reimbursements. Hospitals with larger urban footprints, like HCA or Tenet, are more insulated given their relatively larger concentration on commercial insured providers. Rural hospitals stand to lose a lot, as do behavioral health facilities, like those managed by Acadia. Behavioral health facilities treat mental health, substance, and counseling issues, which are highly prevalent in typically less well-off groups of people. Separately, the Inflation Reduction Act of 2022 allows Medicare (for the elderly) to negotiate prices for certain high-cost drugs. This can mean lower prices on medicines => lower future revenue => reduced pricing power for large pharmaceutical companies.

Industrials This includes aerospace and defense, industrial machinery and equipment, transportation, construction, energy infrastructure, and environmental services.

Reshoring has been a standing theme. For example – GlobalFoundries announced a $16 billion US investment in semis manufacturing and advanced packaging; Stellantis unveiled a $13 billion US manufacturing investment; and J&J plans to invest $55 billion in domestic manufacturing and research facilities over the next several years. Manufacturing construction here in the US also remains near record highs. Every month, new manufacturing construction has been running at $16–20 billion per month. US spending on this has more than doubled since the pandemic. Semiconductors, of course, remain at the center of this reshoring effort. TSMC has committed upwards of $165 billion to its Arizona ops. Intel is continuing to invest many tens of billions of dollars into its domestic fab projects in Ohio and Arizona. And Micron has announced plans that could exceed $100 billion of investment in NY. The key/natural intuition here is that AI demand is reinforcing this push for domestic semiconductor production and supply-chain security.

There are clearly lots of dollars to go around, which means that the bottleneck is increasingly centering on labor. Fabs need enough skilled workers (think technicians, engineers, and semis specialists) – there was a Deloitte and Manufacturing Institute study that estimated 1.9 million manufacturing jobs to stay unfilled by 2033. These job openings remain despite an anticipated slower economic growth. Defense spending, further, is adding an expected 250,000 workers to this labor demand. I think that the engineering and construction firms will be some of the earliest beneficiaries. This includes companies such as Jacobs, AECOM, Fluor and Quanta Services. They will benefit from the heavy invested dollars without requiring factories to first become operational. On these labor shortages, there are a growing number of investments in automation/industrial software. Another bottleneck, if you will, is power infrastructure. Semis fabs and other manufacturing facilities require enormous amounts of electricity. Grid interconnection and transmission issues are beginning to slow some projects, so investing in companies like Vertiv may be a smart play. I will note, though, that most of these reshoring projects have not reached full completion; typically, they have a gap of 3–7 years to move from announcement to meaningful production.

Though I mentioned it above, defense spending is a theme of its own. There are lots of security concerns as AI continues to develop globally, and, as Alex Karp puts it, too many people are focused on the silly problems of optimizing ad spending versus protecting America. We are seeing VC funding flow into defense cos at unprecedented levels – Anduril raising a $5 billion H, Shield AI raising a $2 billion G, and Saronic raising a $1.75 billion D. Lessons from modern conflicts are reinforcing this perceived value in drones, electronic warfare, and low-cost autonomous systems. That’s also why the government is making it easier for venture-backed companies to compete for DoD contracts than the decade prior (for Palantir, this was an arduous process that took about 7 years since the company was founded).

Lastly, automation is increasingly becoming a necessity. There are looming labor shortages (we discussed this), that may be filled thanks to the next wave of industrial automation. Many of the manufacturing hub states (Michigan, North Carolina, Ohio, etc.) strongly feel this labor shortage. The focus for many of the relevant startups (Skild AI and Hadrian being two of them) is designing systems that can learn, adapt, and operate in dynamic environments. The software factory stack is also something of note. Established players like Rockwell Automation, Honeywell, Siemens, and Emerson are embedding AI and software deeper into factory operations – it’s their goal to become the operating system for industrial production rather than simply hardware providers.

Technology This includes software, semis/hardware, internet and consumer, IT services, cybersecurity, and AI infra/models.

The AI infrastructure supercycle is the dominant capex story of this cycle. Hyperscaler investment in AI infra buildouts is estimated to increase by more than 50% this year. This is in light of the demand for AI accelerators including GPUs/custom non-GPU chips. Four of the Mag 7, in fact, have committed $650 billion in 2026 capex for AI infrastructure development (this is a 70% YoY increase). The semiconductor market is a direct beneficiary – global semis revenue is expected to hit $1.3 trillion this year (a 53% increase!) High-bandwidth memory is now the primary supply constraint – most HBM capacity is already committed through 2026 and part of 2027 (mainly concentrated in Nvidia and AMD GPU platforms alongside hyperscaler custom silicon programs). HBMs move more data per second than standard memory, are more efficient for heavy workloads (lower power), and are a shorter distance to the chip (reduces latency/energy loss). SK Hynix, Samsung, and Micron are the main players here.

The GPU monopoly is beginning to crack at the edges. Nvidia’s $20 billion deal to license Groq’s technology on Christmas Eve 2025 (including bringing over much of the team) was a signal that even Nvidia sees a world where GPUs are not the only relevant chips for the next phase of AI, being inference at scale. In time, we saw a broader revaluation of the inference chip startup ecosystem – Cerebras ($40 billion), Etched ($5 billion), and d-Matrix ($2 billion). Tenstorrent is another interesting company, which is pursuing a RISC-V open architecture (“risk five”...lower cost, customization, and innovation-friendly) and a business model across IP licensing, chiplet sales, and complete systems. All these companies underlie a specific thesis that inference workloads will overtake training ones in enterprise settings, so the cost/energy efficiency of GPUs would become a liability.

A less-discussed shift is also happening in power semiconductors. AI semis are power-hungry! A modern AI rack consumes about 1 MW of electricity (the equivalent of powering 1,000 homes). The traditional silicon-based power components don’t really handle this well, which is where two newer materials come in: silicon carbide (SiC) and gallium nitride (GaN). These are wide-bandgap semiconductors, meaning they can handle higher voltages and temperatures with less energy loss. These materials are actually already used in electric vehicles. The market for these materials is at $2–3 billion today and projected to reach $16 billion by 2034. Players like onsemi and Wolfspeed are competing for leadership, but manufacturing these materials at scale (in specific, the crystal growth process) is technically difficult.

A second theme in tech is the decline of many enterprise SaaS businesses, which I have written about extensively prior. You can read my February article on the future of software. My view hasn’t changed much – perhaps I am actually less confident in some of the “adaptation” incumbents I talked about. If I were sitting on the IC for a PE firm broadly investing in tech, I would think about launching growth tech strategies, effectively as a hedge against the presumed 50% of my portcos that are squarely SaaS businesses.

Cybersecurity is a very intriguing exception to this broader software disruption narrative. AI is making it more valuable. It makes sense. As AI tools are more used, hackers and nation-state actors can also use them to write sophisticated malware, launch more convincing phishing attacks, and probe systems for vulnerabilities at unwarranted speeds. For instance, in late May, security researcher Taylor Hornby disclosed a critical Zcash Orchard shielded pool bug that could allow undetectable counterfeit ZEC creation, which he found in 28 hours using Anthropic’s Opus 4.8 model after years of prior human audits missed it (Zcash is a cryptocurrency designed to enable private transactions on a public chain). Companies probably feel an urge to pay for specialized expertise sold by outside vendors regarding cybersecurity. This leads to a very different buyer psychology than in other software categories. 91% of orgs plan to increase threat intelligence spending this year even as software budgets are facing serious pressure. We also are seeing plenty of cybersecurity/threat intelligence/red-teaming startups, particularly across the last few YC batches.

In summary, there are a ton of very interesting things going on, but I’ve only scratched the surface! More to come.

Note: I will probably spend time in another article talking about business services (staffing solutions, outsourcing/BPO, professional services, data services, and marketing services).