188: After Jensen Huang’s Sudden Departure: Nvidia, AI Infrastructure, US-China Competition, and the Compute Ecosystem
Summary
- Todd believes Huang’s departure could weaken Nvidia’s ability to change course, coordinate the organization, and command trust across its supply chain. Several of the company’s transformations over the past 33 years depended on the founder making the call. More than 60 executives report directly to him, with no clear second-in-command or succession plan. Taiwan’s supply chain has invested for years in line with his technology roadmap—“those who listened to Huang made money”—and a professional manager cannot immediately inherit those relationships.
- The bigger industry variable is whether the “AI central bank” shifts from expansion to contraction. Todd likens Huang to an aggressively accommodative central bank governor. Citing Nvidia’s credit support for OpenAI’s roughly $250B, 8 GW buildout plan, he worries that an interim CEO coming from the CFO role could tighten guarantees and financing. Smaller cloud providers such as CoreWeave, which expanded on Nvidia’s credit, could come under pressure first, turning the old “upward spiral” into a mutually reinforcing pullback in financing and orders.
- Todd was bearish on AI infrastructure stocks before this episode; his core view is that capex acceleration has peaked. Nvidia’s 2026 and 2027 revenue estimates he cites remain as high as $400B and $700B, respectively, yet cloud providers are nearing a point where they spend as much as they earn each quarter—or keep borrowing to spend more. Since the second half of 2025, their stocks have made limited headway. In his view, they have entered a phase of “falling valuations, rising and delivered earnings.” He remains bullish on application opportunities over the next 10–20 years.
- Todd reads AMD and Broadcom’s gains as a rotation in market share against a weakening total market. He says AMD’s share of TSMC’s advanced packaging capacity could rise from about 9% in 2026 to 20% in 2027. If Nvidia cuts orders, the freed-up CoWoS capacity could continue flowing to AMD and Google TPU. The “Nvidia stumbles, peers feast” trade may already have played out by the time Huang’s departure was formally announced; the industry’s longer-term capex problem remains unresolved.
- Domestic Chinese chips have a window to replace imports, but near-term growth is still constrained by manufacturing capacity and software compatibility. Liu Yilun says large companies have already booked all available domestic card capacity. Manqi sums it up: “They want to make money from this, but they can’t produce enough.” The bigger question is whether model companies will have to accelerate ecosystem adaptation, moving domestic GPUs beyond a greater role in inference and into large-scale training.
- China’s compute market is seeing a scramble for cards and rising prices, while cloud providers face both appreciation in installed assets and higher costs for new capacity. Todd says his channel contacts are being told, “Buy as many as you can get.” He says B300 server prices rose from around RMB4M before the Spring Festival to around RMB20M in October. The show cites orders from the 3 BAT companies for more than 3 million PRO 5500 cards, worth more than $35B, though actual deliveries could be less than 5% of expectations. Their eventual compliance status is also unresolved.
- Tighter GPU controls could widen the gap between Chinese and US model companies and hit US firms that depend on Chinese open-source models. Chinese model companies may face more restrictions next year on the past route of accessing high-performance compute through overseas cloud providers. Manqi relays one estimate from a researcher at a major Chinese company: for the same workload, OpenAI may have 100 times as much compute available. Further US restrictions on Chinese open-source models could also affect companies such as Fireworks, which uses GLM-5.2 and K3 for private deployments at enterprise customers.
- Todd’s view that the episode is a political purge aimed at Trump is an inference; Manqi leaves room for another explanation. She notes that Republicans still chair the 3 committees seeking Huang’s closed-door testimony on November 18. They may be investigating before Democrats formally take control of Congress on January 3, 2027. From his scenario of cooling AI financing and partisan conflict weakening the dollar, Todd suggests watching long-duration Treasuries and gold, but both trades depend on those specific conditions.
Deep dive
1. The founder’s departure exposes the company’s dependence on him for both strategy and governance
Todd distinguishes the roles of professional managers and founders: the former are good at maintaining strategic stability, but major pivots often need a founder to drive them. Huang has led Nvidia through 4 or 5 major transformations over the past 33 years, most of which look right in retrospect. With someone else in charge, they might never have happened.
Liu Yilun cites several pivotal decisions: pushing CUDA in 2006 when the stock was low and the board opposed it; betting on deep learning in 2013; acquiring Mellanox and pivoting to data centers in 2019; and investing in related companies and supporting the open-source ecosystem after ChatGPT took off. Each decision faced pressure from inside or outside the company.
Nvidia’s other 2 co-founders have long since left day-to-day leadership; one retired in 2003. Liu says Huang directly manages more than 60 executives. The company has no real second-in-command and no succession plan. He worries that without its sole coordinating force, the organization could fracture into “warring fiefdoms,” though the impact may take time to emerge.
Liu says Chinese employees are first worried about the new CEO’s stance on US-China relations and whether the company will cut its China business. But Nvidia already has more than 4,000 employees in China working on a substantial amount of core R&D. Some inside the company therefore believe a sweeping closure of its China R&D center is unlikely.
2. Huang built a personal credit network linking the supply chain to industry financing
In Todd’s view, semiconductors require hundreds of upstream and downstream companies to work together, and Nvidia depends heavily on Taiwanese suppliers. Huang can communicate the next-generation technology roadmap directly, allowing suppliers to upgrade ahead of time before orders arrive. Years of working together have created a shared interest: “Everyone who listened to Jensen Huang made money.”
Todd goes further and calls Huang the “central bank governor of the AI industry,” expanding liquidity and providing credit support to keep the industry loose. He describes the interim CEO, who came from the CFO role, as conservative and cautious. A change at the top could mean a shift in credit policy across the industry.
Todd cites Nvidia’s support for OpenAI’s roughly $250B, 8 GW buildout plan. Nvidia “should be the lender of last resort,” he says; in effect, it is providing credit. He says the new leader previously opposed the arrangement, so succession cannot be viewed as a company personnel change alone.
CoreWeave is the example Todd worries about. Compared with “big banks” like Google and Meta, smaller cloud providers are more like “rural credit cooperatives,” with expansion dependent on Nvidia credit. If liquidity across the AI industry contracts, CoreWeave could feel it first. If OpenAI cannot secure the planned credit, it may not be able to buy as many chips, turning the “upward spiral” into a “downward spiral.”
3. Nvidia already faced three challenges before this episode
Todd first points to Nvidia’s gross margin of about 70%–75%, which concentrates a large share of industry profits in one company. The products are excellent, but “the whole industry is working for him.” That is a commercial advantage and a powerful incentive for everyone else to find alternatives.
His ASIC analogy comes from Bitcoin mining: once the algorithm is fixed, it can be hardwired into a dedicated circuit, bringing costs down to about 1% of a GPU’s. If Transformer architectures and model sizes gradually stabilize, and the difference between 8T and 80T parameters becomes less meaningful, Todd believes dedicated chips will eventually replace GPUs. He just does not know when the industry will reach that boundary.
His view of CUDA is more forceful: however strong it is, it remains software, and AI could make it easier to port software across chips. Liu points out the irony: Nvidia drove AI progress, and AI could now weaken its moat. Just as the company needs its next transformation, the founder best equipped to drive one has left.
4. Earnings cuts and the loss of the founder premium could weigh on valuation at the same time
Manqi says Nvidia shares fell as low as $148.80 on the day Huang announced his departure, then recovered above $160. Todd separates the near-term financial picture from the long-term strategy: the 2026 revenue estimate of about $400B and 2027 estimate of about $700B he cites on the show do not mean those numbers cannot be reassessed.
Todd roughly estimates that 20%–30% of cards went to places they should not have gone, and thinks the true figure may be closer to 30%. If Democrats take power, tighten oversight, and slow US AI data-center construction, he sees a scenario in which revenue could fall 30%–40%. That scenario rests on assumptions about policy and where the cards went.
The longer-term pressure comes from the shareholder base. Some long-term investors have held for years and are willing to wait through a 20% or 25% decline because they “just trust Old Huang” and “he knows more than I do.” Huang’s departure could erase part of that premium. Even if Nvidia’s 2028 valuation may already be below 10x, Todd worries the market could value it even lower.
5. The infrastructure cycle may turn before earnings growth ends
Todd traces the current cycle back to late 2022: slowing inflation growth led markets to expect monetary tightening to turn, while GPT-3.5 persuaded people that better models could be built by adding compute, data, and talent. “It turned scientific problems we couldn’t figure out into defined engineering problems,” setting off infrastructure investment.
Google, Meta, and Amazon were the major spenders over the past few years, but Todd thinks their cash cushions have shrunk considerably. They are approaching a point where they spend as much as they earn each quarter, or borrow more to keep spending. His view is about the capex cycle in public markets; it does not require server construction or company revenues to stop growing immediately.
He calls Nvidia an “old man stock” for its performance since the second half of 2025: earnings keep coming through while valuation keeps falling. Investors most want EPS and valuation to rise together in a “Davis double play.” Now they may get only the former, which is why Todd thinks investors should already have started considering an exit.
6. AMD and Broadcom are capturing relative gains ceded by Nvidia
Todd observes that AMD and Broadcom rise when Nvidia falls. The Philadelphia Semiconductor Index looks little changed on the surface, but major rotation is happening under the hood. He describes these companies as “Nvidia stumbles, peers feast,” framing the trade as a redistribution of market value within the sector.
His specific evidence is advanced packaging: AMD’s share of TSMC’s advanced packaging capacity is about 9% in 2026 and rises to 20% in 2027. If Nvidia cuts orders further, AMD is likely to take most of the freed-up CoWoS capacity, followed by Broadcom and Google TPU.
Manqi asks whether peers would also suffer if total demand declined. Todd acknowledges that the overall market could still shrink; AMD and Broadcom are simply gaining relative to Nvidia as it weakens. He even thinks the short-term trade may have run its course when Huang’s departure was formally announced, with the direction potentially reversing afterward.
7. The opportunity for domestic substitution is real, but supply cannot fill the gap immediately
When Liu Yilun first discusses sentiment around domestic chip stocks, Manqi brings the conversation back to operating conditions and practical realities. Liu’s central point is that large companies have already booked all available domestic card capacity. The shortage is upstream supply, not downstream orders, so domestic vendors may not be able to take Nvidia’s share right away.
Capacity ultimately comes down to manufacturing and equipment players such as SMIC and NAURA. Liu says SMIC has allocated much of its capacity to Huawei, which still ships 700,000–800,000 chips a year. He expects SMIC to be able to manufacture Huawei chips next year, potentially freeing capacity for the “GPU Four Little Dragons” or Cambricon.
Manqi stresses that whether chips can handle large-scale training also depends on the software ecosystem. She says DeepSeek has used TileLang since 2025, describing it as an open-source system initiated by a Peking University team that can adjust and tune operators, offering an adaptation path alongside CUDA and Triton.
Domestic cards have mostly been used for inference, while training still depends heavily on Nvidia’s high-performance GPUs. If external supply is further restricted, model companies will have to accelerate adaptation, and Alibaba may also work more aggressively to support T-Head. Todd reiterates the logic that “blocking is worse than opening the way”: restrictions could give the local ecosystem room to grow.
8. The scramble for cards is underway, with prices and orders reflecting supply anxiety
Todd says contacts in the channel have given him a clear timeline: “There’ll be a crazy GPU stockpiling rush over the next few days.” They are placing GPU orders and buying as many as they can. Manqi says companies are trying to complete partnerships that are still in progress; they will buy whatever cards they can get, even at a premium.
Todd says a B300 server cost about RMB4M before this year’s Spring Festival and had reached about RMB20M by October. He has not yet checked whether prices surged again after the episode. Manqi says Todd also advised her to stock up on an RTX 5090 consumer GPU, which the show priced at about RMB50,000.
On PRO 5500, Liu relays a Reuters report that the 3 BAT companies ordered more than 3 million cards in total, worth more than $35B. They are viewed as the best compute cards China expects to be able to buy, but the prospects of getting them into the country have shifted from relatively clear to uncertain, with no final answer yet.
PRO 5500 cards have a specific use: in a prefill/decode-split setup, they handle the compute-intensive prefill stage rather than fully replacing B300s. Liu says each card costs about RMB85,000–90,000, and the number that actually reaches China may be less than 5% of expectations. There is still a substantial gap between placing an order and having usable compute.
9. Cloud providers’ installed assets are appreciating, but long-term expansion could cost more
Todd describes the impact on cloud providers as “a double-edged fight”: high-performance cards they already own have become scarce, potentially lifting rental prices and benefiting existing assets. But obtaining cards with comparable performance in the future could cost more, possibly much more. He explicitly puts a question mark over the net effect.
Manqi extrapolates to the competitive landscape: Alibaba has T-Head, Baidu has Kunlunxin, and Tencent and ByteDance may be behind Alibaba on related development. Todd thinks companies with cards in hand could have more pricing power, but warns that specialist chip companies may be as strong as cloud providers’ in-house teams. The eventual landscape will depend on output and competition.
Liu says Huawei Ascend’s major customers include ByteDance and DeepSeek, which mainly place direct orders and own the chips rather than renting them through Huawei Cloud. Manqi suggests that if overseas high-end cards become harder to obtain, closer integration between Ascend and Huawei Cloud could also help the latter grow. The discussion leaves this as a long-term possibility.
10. Compute and open-source restrictions could both narrow the runway for Chinese model companies
Todd believes the gap between Chinese and US models could widen if Chinese companies find it harder to obtain GB300 cards for training or inference, unless domestic cards catch up faster. Manqi relays an estimate from a researcher at a major Chinese company: for the same work, OpenAI may have 100 times as much compute available.
Liu says Chinese model companies have previously accessed high-performance compute through large overseas cloud providers, possibly including cards in Southeast Asia. If that route faces tighter restrictions next year, accessing large clusters could become harder or more expensive. He thinks the recent decline in model company stocks may partly reflect this.
Manqi points to forces in the US pushing to restrict Chinese open-source models. DeepSeek has already been excluded from some military and government-related use. Expanding restrictions to US companies and the broader market remains a possible next step. She thinks Huang’s departure could also increase the odds of such restrictions.
The impact would reach US companies, too. Liu says Fireworks uses Chinese open-source models such as GLM-5.2 and K3 for private deployments at large enterprises, meeting customers’ needs for lower costs and control. Others worry about the cultural, social, and ideological assumptions behind model alignment, putting this kind of business at a complicated policy crossroads.
11. The industry may slow in the short term, but it is unclear who benefits in the US-China contest
Todd thinks that over a span of decades, the episode may be just a bump in the road; in the near term, however, the risk of credit contraction should not be underestimated. Manqi offers another perspective: some Americans worry that AI is advancing faster than people’s understanding of its risks, so a modest slowdown could be seen as a good thing, not just a loss.
Musk previously co-signed a call for a 6-month pause in AI research, but it did not have the hoped-for effect. Todd uses the “chain reaction” in Oppenheimer to explain the competition: as long as one participant refuses to stop, the others fear it will keep going in secret. An arms race is hard to end through unilateral restraint.
Manqi also points out that if Democrats limit the US AI industry’s “blind sprint” while imposing stricter controls on GPU sales to China, both sides could lose. Whether the US-China gap widens or China gets a window to catch up will depend on what happens next. Todd’s idea of “Trump’s New Deal failing” is a joking extrapolation from a century-long historical perspective.
For people and companies caught between the US and China, Manqi sees mounting pressure to choose sides. Todd sees Huang, Lisa Su, and Lip-Bu Tan’s links to China and Taiwan’s supply chain as a competitive advantage, while also believing those executives will face challenges. Manqi says even founders running top companies may not be able to walk away unscathed by “standing in the middle and connecting both sides.”
12. Todd sees a political purge; the host leaves room for another reading of the investigation
Todd believes the episode escalated after Trump lost control of both the House and Senate. He infers that while Huang appears to be the target, Trump and his allies may be the ultimate target. His description of a political playbook—first damage someone’s reputation, then remove them from office, then hold them accountable—is a hypothesis, not a behind-the-scenes arrangement established by the show.
Liu adds that the October 2025 New York Times report covered Nvidia’s relationship with MegaSpeed, potential chip sales to restricted regions, and a Commerce Department investigation. It did not report the close relationship now being discussed. He also relays a Wired report that Huang had contacted a firm specializing in scandals involving politicians and celebrities after the article and around the time of the Commerce Department investigation.
The investigation has several strands. The case moved from the Commerce Department to the Justice Department, while the FBI is looking at whether classified information was leaked. After related reporting on October 29, 3 congressional committees asked Huang on November 5 to testify in a closed-door session on November 18. The central issues are Nvidia’s security measures and its relationship with Huang Le.
Liu emphasizes that the 3 committee chairs are still Republicans and political allies of Trump. They may also be investigating within their own sphere of control before Democrats formally take control of Congress on January 3, 2027. Todd expects the investigation to escalate and worries Musk may also be caught up in it. Musk and Palantir’s head have publicly voiced support; the silence from key people at Meta, OpenAI, and Anthropic is discussed as a different response worth watching.
13. Whether Huang can return depends first on the investigation, not on boardroom maneuvering
On personal liability, Todd boils the central questions down to two: “Did he know?” and “Did he agree?” In his scenario, not knowing, explicitly objecting but having subordinates proceed anyway, and knowing and agreeing would create very different liability. He even calls the worst-case scenario “treason,” which is his personal judgment, not an investigative finding.
Manqi says some people are studying the company bylaws to find a route to reappoint Huang to the board and bring him back to the company. Todd thinks that will be “difficult in the near term.” The priority is dealing with the investigation. He stresses that this is just a hypothesis; what happens next remains to be seen.
There is no clear answer on a successor either. Todd relays a joke in a group chat about Cook: Taiwan’s supply chain would get “the strictest father in history” because Apple keeps the profits for itself. Manqi points out that at least he knows Taiwan’s supply chain. Cook and AMD’s head are names mentioned in the discussion; the only leadership style on display so far is the cautious approach of the interim CEO.
14. Compliant products remain available, but H200’s path into China is uncertain on several fronts
In Liu Yilun’s account on the show, as of November 2026, Nvidia’s main products still legal to sell in China include GeForce consumer GPUs, RTX professional workstation GPUs, and some edge-computing platforms. From the A100, launched in 2020, through B300, several high-end products have been subject to export controls. A800 and H20 were previously allowed in through compliant variants with lower bandwidth or performance.
Liu describes H200 as requiring case-by-case licenses. Before shipment, a third party must verify the specifications. The cards cannot displace capacity for US customers or go to the military or restricted entities. China is also weighing the development of domestic chips, leaving the path for H200 unclear there as well.
On arrangements in China, he says, “I’ve heard” one proposal is to designate an area in Yunnan and build a dedicated data center to house the servers, preventing them from flowing to other parts of the mainland. This is a reported idea; the show does not say it has become an implemented policy.
He also relays foreign media reports that Hong Kong has 600,000 H200s, prompting companies to look for local data-center space. To bring them online and monetize them quickly, some are reportedly willing to pay 2–3x the going rate to rent 3–5 MW of data-center capacity. Compared with projects in Ulanqab that run to 1–2 GW, the power footprint is nearly 1,000 times smaller, but locations where high-performance cards can be deployed have more immediate value.
15. Overseas purchases and overseas use may still face scrutiny
Liu Yilun divides the controls into 4 layers. The first is destination: exports of advanced chips above the relevant thresholds to mainland China generally require a license. He says Macau is controlled at the same level as mainland China. On the specific wording for Hong Kong, however, he says, “From what I’ve seen, it looks like it wasn’t included,” without reaching a firm conclusion.
The second layer is the ultimate parent company. He cites a BIS explanation dated May 31, 2026: even if a recipient company is registered in a third country such as Singapore or the Cayman Islands, advanced computing products still require a license if its headquarters or ultimate parent is in mainland China or Macau. Setting up an overseas subsidiary does not automatically get around the restrictions.
The third layer covers the end user and end use. Even if the buyer is entirely foreign-owned, an exporter may still trigger controls if it knows that the chips will ultimately be transferred to a restricted Chinese entity. The fourth is the Foreign Direct Product Rule: some chips made by overseas manufacturers such as TSMC and Samsung using US technology or software may also fall within its scope. Liu says this points to further tightening of controls.
16. Compute demand has shifted from cancellations to a buying spree, reversing the mood at leasing companies
Liu reminds listeners that 2025 was a buyer’s market. Local intelligent-computing centers had capacity they could not sell, and some model companies were even cancelling leases. This year, the idle capacity has been booked. The change comes from both sides: imports of high-end servers face tighter constraints, while demand from coding and agents has surged, including demand for training larger models.
The clearest contrast is among financial leasing companies. Last year they worried that large companies would cancel leases and leave loans unpaid; this year, they are “walking on air” and want to put more money into compute. If Nvidia’s GPU business becomes difficult to sustain, capital and partnerships could shift toward domestic chips, OEMs, or CPU demand driven by agent orchestration.
Manqi asks whether Zhipu and DeepSeek will look back and regret not buying cards more aggressively in 2025. Liu agrees, but reconstructs the conditions at the time: models were smaller, agents and high-reasoning workloads were not yet at today’s scale, and financing did not support renting large amounts of compute. The discussion also mentions constraints at the point of going public. Companies did not have today’s information then.
17. Which workloads get cards first depends on internal businesses and revenue targets
Liu believes ByteDance and Alibaba were more aggressive than Tencent about stockpiling cards early on, and spent more overall. Tencent was especially focused on procurement compliance and bought many H20s when they were available for compliant purchase. One detail: others expressed total compute in H100 equivalents, while Tencent internally used H20 equivalents.
The order of allocation also differs. According to Liu, ByteDance prioritizes consumer products, then other internal R&D workloads, with its external cloud platform last. Tencent prioritizes internal businesses such as WeCom, Bot, and SME products before offering cloud services externally. He says Tencent’s external cloud service currently can use only Kunlunxin and Hygon cards.
Alibaba is a special case in Liu’s view. Market attention to Alibaba Cloud’s results pushes it to rent more compute to independent model companies such as Kimi and Zhipu. One reason the cloud team gives is that there cannot be only one surviving model provider in China, so it also needs to back customers beyond Qwen.
Manqi focuses on the internal conflict: Qwen does not have an internal settlement mechanism for using Alibaba Cloud compute, leaving the sales team with little incentive to prioritize cards for Qwen. Alibaba has to balance whether to put more resources behind its own model or serve external customers who can contribute revenue immediately.
18. Token revenue sharing lets independent model companies bring external compute back into official channels
Independent model providers are generally short of cards and lack the purchasing power of large companies. Liu lists their main options: partner with leading cloud providers, work with independent smaller clouds—as in Zhipu’s partnership with Parallel Technology—or explore token revenue sharing to reduce the burden of owning and sourcing compute.
The example of Chuangxin Data shows how the arrangement works. Once tokens produced by the compute provider meet Zhipu’s quality and performance standards, they are sold through Zhipu’s official channels. The price may be higher than selling independently after deploying an open-source model. Zhipu avoids the compute-purchasing cost while bringing commercial revenue that might otherwise flow entirely to a third party back into its own system.
Liu says a common split in current discussions is 70% for the compute provider and 30% for the model provider. Manqi adds that Zhipu’s 30% revenue share with AWS is public information and another model company is discussing a similar arrangement with a major cloud provider. She has not independently verified whether other domestic deals have closed or whether they will also settle at 30%.
19. Todd uses long bonds and gold to express the risks of a cooling AI cycle and US political conflict
Todd’s case for 10-year and longer-dated Treasuries starts with the premise that if rising yields were driven mainly by hot AI financing demand, and that demand has reached its peak, long-term rates could fall. That would support a long-duration bond trade. He cites yields of “a little over 5 points” at the time and considers the cushion attractive; in effect, the trade is a bet on cooling or even recession.
His case for gold rests on a different premise: worsening US partisan conflict could damage the country’s interests and strength, further weakening the dollar. He describes the relationship between the dollar index and gold as “100% negatively correlated,” and favors gold on that basis. This is an allocation view tied to his political scenario; the show provides no further data to support it.
20. Applications still have long-term potential, but high valuations need a second growth curve
Todd thinks infrastructure now accounts for a large share of the AI industry’s revenue and market value, while applications remain small. The industry does not yet look mature. In mature industries, revenue and market value tend to rise the closer a company is to the end user. So his bearishness on the infrastructure stock cycle does not stop him from expecting huge application companies to emerge 10–20 years from now.
That outlook depends on 2 conditions: model capability keeps improving and usage costs keep falling. If both continue, applications that are hard to imagine today could emerge. But he admits that he does not know what they will be: “We’re still very early on applications.”
Asked whether he would buy OpenAI at an IPO, Todd answers “No, no, no—too expensive” at a valuation of about $1.4T. He definitely would not buy at $2T. If the stock spikes on the open, he might consider shorting it. At $1.4T, though, even a short is “hard to say”; being unwilling to buy does not automatically make it an obvious short.
His hesitation comes down to the need for a second growth curve. From 2023 to 2026, after 4 years of effort, coding remains the core use case with genuine positive revenue feedback. Drug discovery can accelerate molecular and target discovery, but cannot bypass Phase I, II, and III trials or the physical constraints of patients and hospitals. Robotics is another large opportunity, but the path and pace are still immature and may have to wait for the next cycle.