2026.10.09 | Substack | Semis | TMTB Podcast Intel: Qualcomm & SemiAnalysis
原标题：TMTB Podcast Intel: Qualcomm’s (QCOM) CEO on AI Phones, SemiAnalysis on the AI Buildout, FCC Chair on SpaceX (SPCX…
TMTB Podcast Intel is produced by Ethmos, research agents that listen to every episode in TMTB’s coverage universe and write the briefs below.
Try Ethmos — 14 days free, then 25% off your first month for TMTB readers.
Today’s Essential Episodes:
- Qualcomm CEO Cristiano Amon on the Coming AI Phone Supercycle, Smart Glasses as the Biggest AI Wearable and the
Modular Bet Against CUDA Sources with Alex Heath, with Alex Heath and Cristiano Amon, president and CEO of Qualcomm
- SemiAnalysis’s Daniel Nishball and Jordan Nanos on Who Funds the $11 Trillion AI Buildout, NVIDIA’s $497B of
Off-Balance-Sheet Commitments and Rising H100 Rents SemiAnalysis Weekly, with Daniel Nishball and Jordan Nanos of SemiAnalysis
- FCC Chairman Brendan Carr on SpaceX’s Spectrum Push Against the Carriers and Disney’s Fight Over ABC’s Licenses *CNBC’s Squawk on the Street, with Carl Quintanilla, Leslie Picker and Brendan Carr, chairman of the FCC (the Carr
segment)*
- Former Intel CEO Pat Gelsinger on Why Manufacturing, Memory and Power, Not Design, Will Bottleneck AI Chips *The a16z Show, with a16z’s Raghu Raghuram and Guido Appenzeller and Pat Gelsinger, general partner at Playground
Global and former CEO of Intel*
- Datadog Chief Product Officer Yanbing Li on Observing AI Agents, Machine-Scale Rate Limits and the Unsolved
Tokenomics of AI The Stack Overflow Podcast, with Ryan Donovan and Yanbing Li, Chief Product Officer of Datadog
Produced by Ethmos: AI-assisted summaries of public podcast episodes. For information only. May miss context or contain errors. Verify before relying on them. Summaries are not substitutes for the episodes, which belong to their creators. Not investment advice, offer, or solicitation. Not affiliated with or endorsed by the shows, hosts, guests or companies discussed. Corrections and publisher opt-outs: support@ethmos.ai.
Listen · Watch
THE SIGNAL
- Amon says memory prices up 5-6x are cutting smartphone units about 20% this year, mostly mid/low tier while premium holds; he calls it supply-driven pent-up demand, not weak demand (our read: favors premium-skewed Snapdragon mix).
- Unnamed frontier AI companies are asking Qualcomm for phones that run 100B+ parameter models, always on, by 2028; some are building phones themselves (ByteDance’s Doubao phone, made by Nubia), implying big on-device compute/memory content gains (our read).
- Amon says Qualcomm is expanding into several new markets at once on roughly flat opex, crediting AI productivity gains in chip development, which supports operating leverage as data center, PC and auto scale.
- Amon’s bar for smart glasses to scale: about 35 grams or less, 6-10 hours of battery, and fashion-grade design, giving a concrete checklist for judging design wins.
01 US-China AI sentiment divide
Amon contrasts Chinese consumer AI adoption, which he says is moving fast into devices, with the US conversation, which he says is dominated by frontier-model development and enterprise adoption alongside notably negative public sentiment. He attributes US anxiety to concerns about job displacement, data-center electricity costs, and AI productivity gains — concerns he says are largely absent in China, where AI is treated more as a natural technological evolution and where he sees heavier focus on physical AI and edge applications. He uses the analogy of the early internet (AltaVista, MapQuest, Orkut) to argue it’s too early to predict AI’s ultimate winners.
02 Guardrails and geopolitics
On AI-pacing debates, Amon — prefacing his views as personal opinion — argues the right answer sits somewhere in the middle: continued investment in frontier development alongside serious guardrails, particularly around cybersecurity, since AI models could let non-specialists find software vulnerabilities at scale. He says Qualcomm builds hardware-level controls (encryption, personal data-graph storage options) to support this. Discussing his May trip to China with President Trump, Amon says semiconductors have become a geopolitical industry and describes the conversations as emphasizing stability through continued commercial ties between American and Chinese enterprises.
03 Smartphone market and the AI-phone definition
Amon states the smartphone market, still smaller than its 2019 pre-pandemic peak, sells about 1.2 billion units annually but is maturing into a replacement-driven market — except for a premium tier that has kept expanding as a share of the market even during contraction, which he credits to users shifting more workloads (previously laptop-bound) onto phones. He defines what he calls an AI smartphone not as bundled first-party chatbot features but as a device where a third-party agent ecosystem — similar to how app stores defined smartphones post-iPhone — can access personal context, memory, and device sensors to take autonomous action, citing the OpenClaw product launch as the inflection point for this vision. He rejects the idea that cloud-based agent products (like Muse) make phones less central, arguing cloud and edge are complementary rather than competing, pointing, from memory (he cautioned he might have the numbers wrong), to NVIDIA’s last earnings as showing roughly 40% of revenue from on-prem, neocloud and sovereign AI deployments as evidence hybrid AI is already mainstream.
04 Wearables: glasses over VR
Amon describes VR as having matured into a console-type, gaming-focused niche rather than a mass computing platform, with mixed reality as a middle step. He identifies smart glasses as the wearable category he is most bullish on, putting the market, by his own estimate, at 50-60 million units shipped (he did not say over what period or exactly which devices that counts) across roughly 40 designs, with GenAI features (translation, object/person identification, see-what-I-see) driving utility beyond the original camera-centric pitch. He expects watches, pins, pendants, and jewelry to also find roles, and likens the broader AI-wearable moment to the cultural inflection of the iPod or Walkman, without naming a definitive third form factor.
05 Data-center entry and the Modular bet
Amon explains Qualcomm’s data-center push as a natural extension of its mobile-derived, battery and energy-efficiency DNA, framing the opportunity around the tradeoff between compute needs and available energy rather than competing head-on with NVIDIA GPUs for training. He notes inference workloads are fragmenting (pre-fill vs. decode) creating openings for specialized, energy-efficient CPUs and accelerators. The Modular acquisition (Chris Lattner, inventor of Swift) is positioned as building an open, hardware-agnostic AI software layer — analogous to Android or Linux — to break NVIDIA/CUDA’s software lock-in; Qualcomm says it will support deployment on competitors’ chips since it only competes on hardware once the software layer is open.
06 6G as AI infrastructure
Amon reviews prior “G” cycles (5G enabling unlimited data and HD video downloads, 4G enabling smartphone-grade broadband) before describing 6G as designed around AI and physical-AI use cases: high-definition video upstreaming from glasses, and treating cellular radio signals as radar-like sensors for real-time mapping, drone detection, and tracking of vehicles and pedestrians across the network. He says 6G networks will carry AI tokens the way earlier networks carried voice and data, enabling distributed compute. Qualcomm plans demonstrations at the 2028 LA Olympics, followed by prototypes, silicon and early deployments, with broader scale after 2030.
Listen
THE SIGNAL
- SemiAnalysis tallies $497B of off-balance-sheet NVIDIA commitments and estimates its balance sheet can backstop only about 46GW, against roughly 240GW coming, so lenders will eventually have to take neocloud risk without NVIDIA’s guarantee.
- One-year H100 contracts are back near $3/hr, close to 1H23 levels, and sellers who weighed $1.30 last fall have renewed at $2.50; this upends straight-line residual-value assumptions for GPU-backed credit (our read).
- SemiAnalysis models frontier-token serving on a GB300 at up to about $100M/MW of annual revenue, against roughly $40M/MW of upfront build cost and about $12M/MW a year in typical neocloud rent ($48M/MW for SpaceX/Google), so labs, not landlords, keep most of the margin.
- Buyers now commit about six months ahead and prepay 30-50% of contract value, which in effect funds the provider’s cost of capital and pushes labs toward self-build, while server OEMs remain stuck near 9% margins.
01 The $11 trillion wall
Daniel Nishball says SemiAnalysis’s capex model sums total cost of ownership — chips, servers, networking, CPUs — across GPUs, XPUs, and TPUs (including AWS Trainium) to arrive at $11.3T of cumulative AI IT and data-center capex from 2024 through 2029, a figure much larger than headline “hyperscaler capex” because it also includes spending by independent data-center operators and neoclouds. From this, the firm derives $7T of AI debt outstanding by 2029, with roughly $1T in net new funding this year — tracking close to the ~$500-600B raised through June. Since June 1 alone, Nishball counts roughly $400B of loans and bonds either raised or announced, including a $24B Blue Sky Funding deal tied to the Broadcom/Google/Anthropic TPU partnership, a $12.5B Meta-BlackRock JV for a data center in El Paso, Hut 8’s $4.25B raise, CoreWeave’s $4.2B convert plus $2.6B delayed-draw term loan, Nebius’s $3.5B convert, and Anthropic’s $15B pre-IPO revolver.
02 Constraints have shifted
The bottleneck has moved over time, per Nishball: memory shortages previously forced chip “despec” (e.g., reduced SOCAMM on NVL72 racks) to preserve shippability; data-center siting was solved through creative capital deployment, powered-site activation, and behind-the-meter builds; now financing is the acute constraint. Nishball notes data-center capex is understated relative to chip capex on a cost-of-ownership basis because data centers depreciate over 15 years versus 5 for GPUs — meaning one data center hosts roughly three cycles of chip capex.
03 Backstops and balance-sheet-as-a-service
NVIDIA’s seven backstop categories — including AI cloud agreements, data-center leases, guarantees/LPs, cloud services agreements, capacity commitments, and residual value guarantees — exist, Nishball argues, because deals require a five-year offtake to get financing at all (the “Trinity” of capital, offtake, and data center), and most offtakes run through hyperscaler guarantees (e.g., Microsoft backstopping neocloud sales to OpenAI). This keeps commitments off balance sheet under accounting treatments like ASC 606 for managed services versus operating leases. Amazon’s $25B, six-tranche bond priced wide in July as an early warning sign, per Nishball; Amazon’s 10-year Z-spread moved roughly 10-20bps wider post-issuance, and Oracle’s curve has widened ~30-50bps since summer, with Oracle’s current 10-year Z-spread around 250bps versus Amazon’s ~100-110bps.
04 ClusterMAX as a credit signal
Nishball frames CoreWeave’s lending spread (roughly 200bps versus Meta’s ~97bps curve) as reflecting execution risk on top of the offtake counterparty’s credit, suggesting ClusterMAX ratings — SemiAnalysis’s own cluster-quality grading — could proxy for that execution risk in lender roadshows. Nanos pushes back that managed clusters and bare-metal products are different animals and that site-level SPV structuring may be a better unit of credit analysis than company-wide ratings.
05 Token profitability and the demand case
Nanos argues pricing for H100s is rising less because of model-training preference and more because GB300 and B200/B300 capacity is sold out, pushing demand down-market; Chinese model efficiency gains (DeepSeek v4, GLM 5.3, Kimi K3) show 80% gross margins on GB300 NVL72 serving even for open models. SemiAnalysis’s income-statement model for a frontier lab puts blended revenue at ~$1.85 per million tokens (~$38/hour equivalent) against a $2.80/hour cost of ownership, implying roughly 90% API gross margins, while all-you-can-eat consumption plans are murkier depending on usage intensity. Both hosts argue this profitability — alongside personal-productivity gains from AI agents, which Nanos says he would never want to give up once experienced — supports continued demand independent of further model improvement.
06 Structuring the backstops
Nishball details specific backstop mechanics: NVIDIA’s AICP program (started with Sharon AI, then Firmus, with GMI and others following) guarantees it will rent unused neocloud capacity for six years, letting lenders price risk as “NVIDIA plus execution” rather than requiring investment-grade offtake; NVIDIA also takes a 50% revenue share above guaranteed minimums, and uses similar structures with Lambda and a roughly $105B residual-value guarantee behind one large data-center complex. A proposed $500B capital partnership structure would tranche risk (equity, B-piece, A1/A2 senior) with a 25% NVIDIA residual-value guarantee attached to specific tranches, opening the market to different investor risk appetites — banks, insurers, private equity, private credit. The newly formed Compute Capital Markets group is building a public-domain database of thousands of debt facilities to support this analysis and is actively hiring credit analysts in New York and Singapore.
Listen · Watch
THE SIGNAL
- Carr said FCC spectrum deal flow already tops $100B, and auctions should bring a large amount of additional spectrum to market in about two years, an overhang that could cap spectrum valuations (our read).
- An auction of 25 MHz, which Carr placed between low and mid band, is coming soon, giving SpaceX, Amazon or the carriers another way in and making SpaceX’s new low-band holdings less scarce (our read).
- Carr said carriers are moving away from high band and that any serious terrestrial player needs both mid and low band, which SpaceX now holds, so the chairman is effectively calling its portfolio network-capable (our read).
01 SpaceX’s spectrum land grab
SpaceX’s purchase of a national block of low-band spectrum, announced the day before the interview, sent AT&T, Verizon, and T-Mobile shares lower. Carr explains the significance using FCC spectrum taxonomy: low-band spectrum covers wide geographies and penetrates buildings, high-band offers high capacity but limited range, and mid-band is the compromise he calls the “Goldilocks” option. SpaceX already held mid-band spectrum from its earlier EchoStar deal; the new low-band acquisition, Carr says, signals that SpaceX is serious about building a genuine terrestrial wireless network rather than remaining a pure satellite play. He notes there’s disagreement in the market over whether SpaceX actually intends to build traditional towers, and says the FCC is simply watching rather than adjudicating that question.
02 FCC’s hands-off, pro-deal posture
Carr repeatedly stresses that the FCC’s role is not to pick winners and losers — that market forces and consumer choice will determine outcomes among SpaceX, Amazon (which he says is launching its own direct-to-device satellites later this year), and the traditional carriers’ joint venture on direct-to-cell technology. He frames the current wave of consolidation as a direct result of EchoStar’s prior exit from mobile wireless, which pushed spectrum back into circulation. Carr contrasts this with what he describes as a regulatory “wet blanket” under the Biden administration, under which sector deal flow fell to roughly nothing and spectrum sat underused. He cites a recent study showing high-speed fiber internet prices falling 7-14% for consumers over the past year as evidence that competition from spectrum rationalization is already benefiting consumers, tying it back to CPI relief.
03 Carrier pushback
Anchors relay Verizon’s statement that satellite “complements” rather than replaces terrestrial networks, and that acquiring spectrum alone doesn’t guarantee the infrastructure needed to compete as a true nationwide mobile provider. Carr does not directly rebut this, instead reiterating that the FCC will monitor how the market plays out and let competitive dynamics — not regulatory intervention — determine whether SpaceX’s bet pays off. An upcoming FCC vote later this month will decide whether to proceed with a new spectrum auction for direct-to-device services, the same underlying technology used by Starlink.
04 Disney’s ABC license fight
Carr characterizes Disney’s attempt to block the FCC’s early review of licenses for its eight ABC-owned TV stations as a procedurally flawed end run — arguing under the Communications Act that any objection to an FCC renewal decision must go through the full commission and then the court of appeals, not directly to a federal judge. He says a DOJ lawyer representing the FCC argued Disney filed in the wrong venue at the wrong time, and expresses confidence the FCC will prevail on that threshold issue alone. Separately, he says the FCC has a significant investigation into Disney under way, notes that petitions to deny filed against Disney raise potential DEI-related discrimination and its other conduct during its license term, and says the agency has a dozen or more proceedings open looking at broadcasters generally — not an isolated focus on Disney.
05 Public interest standard and press trust
Asked whether Disney’s news division softened its coverage of the administration under regulatory pressure, Carr argues Disney’s courtroom admissions (reduced airtime for Democratic candidates, altered operations) contradict its public messaging about standing firm against the FCC — implying that if conduct changed only once scrutiny began, it wasn’t compliant before. He says the FCC’s public-interest standard centers on localism — broadcasters identifying and serving local community needs — and separately argues that declining public trust in legacy broadcast journalism, per Pew and other studies, is bad both for the industry and the country. Carr says no decision, or timeline for one, has been made on the ABC licenses themselves.
Listen · Watch
THE SIGNAL
- Gelsinger called a recent Oracle comment the first of many signs that energy shortfalls will cause defaults on data center projects, and expects buyers to check that power can underwrite their GPU and rack commitments.
- He expects incumbents to absorb heterogeneity: Groq will be indistinguishable inside NVIDIA within two years, and OpenAI, NVIDIA and Anthropic will pick which of about 100 chip startups survive (our read: big-customer traction is key).
- Copper scale-up now costs more at 5 meters than optics at 100 meters, but CPO was held back by thermals, packaging and too little laser capacity (our read: watch laser supply before 2028-29).
- He doubts optically disaggregated memory (moving bits costs about 1,000x compute energy), processing-in-memory, and flash or MRAM approaches; he favors stackable ferroelectric-type materials and has funded a stealth memory startup.
01 Design speed vs. manufacturing drag
Gelsinger frames the core tension: AI is solving the “easy” part of chip design — logic synthesis, RTL generation, kernel writing — while analog components like SerDes remain resistant to AI automation and still require extensive silicon validation. Even with design compressed to three months, he says fabrication, advanced 3D packaging, and rack-scale integration still stretch the timeline many months further before a chip can be used — long enough that a chip’s architecture may no longer match the AI workloads it was designed for. He points to Graphcore historically as a case where a reasonable chip design was overtaken by workload evolution before it could scale, and frames compressing this manufacturing/packaging timeline (not design) as the critical innovation challenge: new lithography approaches and cheaper prototyping flows that don’t require $50 million in mask costs before testing.
02 Memory and power as physical limits
Beyond manufacturing lag, Gelsinger names memory bandwidth and power delivery as parallel bottlenecks. He calls HBM “a hideous memory” — poor bit density, shoreline bandwidth limits, thermal issues from concentrating heat — despite being the best current option, and sees compute-memory stacking (citing portfolio company d-Matrix, and referencing Cerebras) as the likely path forward, though he expects stacks to plateau around two to four memory layers (not 16+) due to exponentially worsening yield requirements at higher stack counts. On power, he says most power simulation and three-dimensional thermal modeling tools remain immature, forcing roughly 40% power loss to guard-banding; he’s bullish on 800V DC delivery, solid-state transformers, and vertical GaN conversion technology, and cites a portfolio company, Snowcap, pursuing superconducting approaches that could offer roughly 1,000x better power performance, addressing CMOS’s flat power-per-teraflop across the last five NVIDIA chip generations.
03 Chip architecture consolidation debate
Raghuram and Appenzeller push back on Gelsinger’s consolidation thesis, noting AI-driven design tools may make heterogeneity cheaper to sustain (easier to write optimized kernels for diverse hardware) rather than forcing standardization as in past eras. Gelsinger counters with three reasons he still expects convergence: workloads are becoming more granular (pre-fill, mid-fill, decode specialization) in ways he calls historically unsustainable; LLM architectures may hit algorithmic limits requiring different compute profiles (e.g., high-precision 64-bit math for chemistry/molecular modeling); and capital-market dynamics will naturally eliminate most of the ~100 current AI chip vendors since achieving scale requires winning workloads and capital. He also predicts training and inference will converge toward continuous learning rather than remaining separate, undercutting the case for permanently specialized architectures — though he concedes workload-specific optimization (e.g., for diffusion models) could still justify some fleet heterogeneity.
04 Optical interconnects and networking shift
Gelsinger reiterates a long-held view — he says he declared copper dead about 25 years ago — arguing all I/O should move to optical, while core compute-memory complexes should not, due to prohibitive power costs of optical-electrical-optical conversions. He describes NVIDIA’s NVL72 as a remarkable engineering achievement that was nonetheless brutally difficult to manufacture, noting the industry needed roughly 18 extra months to absorb it, and argues a more mature optical supply chain would have enabled better scale-up systems sooner. He expects co-packaged optics (CPO/NPO) to reach capacity around 2028-2029, and anticipates AI networking will shift toward circuit-switched/optical-switched designs given AI’s large, predictable traffic flows — contrasting with traditional packet networks built for unpredictable routing, an idea he attributes to Stanford’s Nick McKeown.
05 Energy capacity and the VM-for-agents close
Gelsinger stresses US energy capacity has structurally lagged, with renewables offsetting retiring coal roughly 1:1 for over a decade before accelerating to about 4% annual growth over the last five years — still insufficient, in his view, to support AI buildout ambitions, raising the risk of capital-intensive data center projects defaulting on power commitments. He also criticizes dependence on China for renewable supply chains and notes gas turbine lead times now run about eight years. The conversation closes on Gelsinger’s VMware legacy: he argues the virtualization abstraction remains foundational but must be rebuilt for a world where AI agents, not humans, are the primary users — covering agent security, performance management, migration (“v-motion for agents”), and human-set policy layers akin to Anthropic’s “constitution” concept, while acknowledging this is a largely untested design space.
Listen
THE SIGNAL
- One customer ran Datadog’s SRE investigation agent over a few weeks, averaging just over three minutes to root cause with no human and no dashboards opened, which is concrete evidence for paid AI-SRE adoption.
- Li says customers keep asking Datadog to raise rate limits on its own platform, implying agents now query Datadog APIs at machine scale; that could open a usage-based monetization path or create cost pressure (our read).
- Li frames incident war rooms of tens to hundreds of people as mostly teams proving their innocence; investigation agents remove that labor, which supports a labor-replacement pricing story rather than a seat-based one (our read).
- Datadog builds its own models, and Li says a lot of AI’s lift to observability demand still comes from core metrics, traces and logs, so agents should boost the base business, not just new SKUs.
01 From self-driving trucks to AI agents
Li’s path runs through Google (leading observability for Cloud and internal infrastructure), then Aurora, where she helped build commercially operating driverless trucks — she notes Aurora remains essentially the only company running driverless trucks on US public roads. That experience, she says, taught her what’s required to “fully unleash your agent into the wild,” drawing a direct parallel between L4 autonomous vehicle validation (heavy offline testing plus continuous closed-loop monitoring once deployed) and how LLM-based agents now need to be developed and observed.
02 What’s new versus what’s familiar
Li states that AI agents are still fundamentally distributed systems, so the traditional observability pillars — logs, metrics, traces, events — remain foundational and demand is growing for both the old and new layers simultaneously. What’s new is a layer of model-specific signals: prompts, responses, reasoning, tool calls, full agent trajectories, and token usage. She cites a concrete pattern from Datadog’s own agent-observability product data: a large chunk of agent errors trace back to hitting API rate limits, an old problem that becomes more pronounced because agents access data at much higher speed and scale, driving customer demand to raise rate limits.
03 Closing the build-production loop
Li frames the central shift as moving from simply observing production health to running continuous evaluation (including LLM-as-a-judge techniques) both in production and during development, then feeding findings back upstream to improve the agent. Because real-world behavior can’t be fully captured by pre-release testing, she argues this feedback loop must be faster and more constant than in traditional software, extending into automating parts of the SDLC itself — autonomous test generation, feature-flagged rollouts that expand or roll back based on signals, and parallelized agent-driven PR review and testing. Datadog is experimenting with this internally and building local-agent tooling (via Datadog MCP) so developers can access production context directly from their laptops.
04 Blurring developer/SRE boundaries
Li says AI is dissolving the traditional knowledge boundary that separated development from operations: SREs can now investigate and potentially patch code they don’t own, while developers increasingly consume production telemetry directly through MCP during development. She frames this as evidence that the SDLC itself “is being rewritten.” On whether LLM evaluation belongs inside an observability product category, Li declines to draw a strict categorical line, arguing instead that engineers building agents benefit from build-time and production-time evaluation tools being tightly connected, since agent observability isn’t an isolated capability but part of a broader system/service observability story.
05 Consumption surfaces and dashboards’ staying power
On how much context to surface via MCP, Li says Datadog has had to curate formatting and query mechanics rather than ship a raw API wrapper, and notes an emerging “interesting battle” over whether customers consume telemetry via lighter MCP access, Datadog’s own curated AI agent, or third-party agents synthesizing data from multiple vendors. She argues dashboards aren’t disappearing — some customers still find clicking through known dashboards faster than natural-language queries — but consumption surfaces are proliferating (Slack, CLI, IDE), and AI agents increasingly generate dashboards just-in-time for a specific investigation rather than reading pre-built ones.
06 Unsolved frontiers: security and tokenomics
Li identifies deep agent-behavior understanding as only “the first step,” with trust and security as the next layer — she calls AI security unresolved and “multi-edged,” since AI can be the target, the attacker, or the attack vector, and argues security and observability increasingly overlap because both require understanding every agent action. She also flags tokenomics/ROI as an open problem: unlike cloud infrastructure, where footprint growth tracked business growth, AI’s value-to-cost relationship is not directly correlated, and the industry hasn’t yet linked token-level cost efficiency to business outcomes. Datadog, she adds, already performs both static and dynamic model routing internally, mixing large models, small models, and basic ML models depending on reliability needs.
POWERED BY Ethmos for TMTB
Get briefs like these for every company you follow.
TMTB Podcast Intel picks a handful of episodes a day. With Ethmos, research agents listen to every podcast that touches the names in your book and file briefs like these every morning.
+ Qualcomm · + NVIDIA · + SpaceX · + Intel · + Datadog · + Your own list
Start your 14-day free trial
TMTB readers get 25% off their first month. Cancel anytime.