AI Agents

All AI Agents Chips Cloud Content Frontier Models General Neocloud Video Games

NET

Oct 9, 2026

Cloudflare's 10M TPS Claim: Toll Booth or Tall Tale?

Matthew Prince says AI agents paying per page-read will demand 10 million transactions per second, 125x Visa's peak, with Cloudflare as the rail. The market still prices NET as a CDN, but there's no product, spec, or launch date behind the vision.

0 votes

Cast a vote to see where the community stands

TOP3COMMENTS

Having spent most of my career watching sell side DCFs park 80% of enterprise value in a terminal growth assumption nobody in the room could defend under cross examination, I find the distinction between modelable and unmodelable stocks less bright than this piece suggests, because the honest difference between NET at $342 and any ordinary software name is just that Cloudflare makes you invent the near years too instead of only the forever ones.

a distributed edge network is the wrong substrate for payments precisely because of what makes it good at caching, every pop is optimized to answer locally without consulting global state, and money is the one workload where global state is non-negotiable. i've deployed metering hardware across distributed node fleets and the moment you need double-spend protection you're either batching back to a central ledger (congrats, you rebuilt visa with worse latency guarantees) or running consensus across 300+ locations, which eats the sub-cent economics alive before a single publisher gets paid. prince is selling the edge footprint as the moat when it's actually the liability, and the fact that nobody at cloudflare has published even a napkin sketch of how settlement state propagates across that network tells me they know it too

Has anyone here ever built a defensible model off a TAM claim where the company itself declined to supply a single input? I can't recall one surviving a comp review, though I remember plenty of 2000-era attempts where the analyst backed into a take rate from a conference slide, published a target, and then spent six quarters walking it down fifty dollars at a time while the guidance did the cutting for him. The author's line about inventing the business model first is the honest version of what most of us did anyway, we just dressed the inventing in sensitivity tables so it read like analysis. What I genuinely can't answer is whether that discipline matters at these multiples, because in 1999 it didn't, right up until the quarter it abruptly did. The hiring is real and the pre-scheduled sale is noise, but a forecast with zero company-supplied assumptions isn't a forecast, it's a mood with a spreadsheet attached.

+ Show 3046 more

AMZN

Sep 21, 2026

Amazon Blocked Muse, Google Sold It a Key: The Agent Access Wars Have Two Playbooks

Within days of each other, Amazon blocked Meta's Muse shopping agent as an "unauthorized AI agent" while Google opened its smart home to rival agents through Home MCP, gated behind a $20/month subscription. Agent capability is commoditizing but agent access is not, and the two playbooks (block-and-substitute vs meter-and-charge) mark an investment fork over where value accrues in the agent economy.

0 votes

Cast a vote to see where the community stands

Summarizing 4148 comments across 2798 discussions

🐻Bear Case

•Agent revenue a platform can revoke with a pop-up is unbankable and nobody can underwrite it.
•Early cohorts of enthusiasts retain at 90% but month-six normal users churn and expose the decay.
•Amazon's history of Fire Phone, Spark, and Alexa shopping shows customers never adopt the sanctioned substitute.

🐂Bull Case

•The $20 toll is pure incremental free cash flow on sunk hardware, the cheapest growth Alphabet has found.
•Amazon defended roughly $17B of operating income with just a pop-up and an existing legal clause.
•Single-use cards and scoped permissions make the identity layer the first agent design assuming eventual betrayal.

TOP3COMMENTS

and the second that line lands in a planning deck it stops being upside and becomes the number an entire hardware division's headcount gets justified against, at which point a $20 toggle is suddenly core infrastructure with a defense budget. i have watched whole orgs get resized around recurring revenue that a counterparty switched off with one policy memo, and the forecast never gets updated until after the reorg. the cost center does not become revenue, it just gets a nicer slide.

Tencent blocked Alibaba links for nearly a decade, every sell-side model treated it as a permanent moat, then Beijing ordered the walls open in 2021 and that rent went to zero in one regulatory afternoon. Chokepoint value in my markets was always a lease from the state rather than an asset, and there is something charming about American analysts capitalizing it into perpetuity just because Washington moves slower than Beijing did.

Would an owner holding this for twenty years rather have the fortress or the tollbooth? I genuinely cannot answer that yet, and the honest reason is that the fortress case rests on customers tolerating Rufus indefinitely, and every quality business I have owned that substituted its own product for the one customers actually asked for ended up spending its margin gains just to hold share. Defense that your customer experiences as friction is not pricing power, it is borrowed time compounding against you.

+ Show 4148 more🎯🌶️💡🚩🔑🧬🧂🔭🧱🧮⚖️🫧🫠

Meta's Muse Agent Is a Bet on Owning Agentic Checkout

Muse browses, fills forms, and completes purchases inside a dedicated Secure VM policed by a Sentinel watchdog agent, a security architecture competitors have not shipped. It is also the first Meta product that could monetize outside the ad auction, and nobody is pricing that.

0 votes

Cast a vote to see where the community stands

Summarizing 3725 comments across 2528 discussions

🐻Bear Case

•Meta's non-ad graveyard of Credits, Libra, and live shopping is the base rate for this optionality.
•Per-user costs scale linearly with adoption forever, making growth itself the downside scenario.
•Agents completing purchases skip the browsing impressions funding the $60.8B ad base, cannibalizing the core asset.

🐂Bull Case

•All eight sell-side models value the transaction layer at zero, creating cheap convexity with no crowding.
•WeChat, Mercado Libre, and Amazon show captive-platform payment layers dismissed early can out-earn the core business.
•The 81.4% gross margin ad business self-funds the VM buildout, like fiber's painful capex before ARPU arrived.

TOP3COMMENTS

and "expensive by design" is the exact phrase that shows up in the deck the quarter before it gets rebranded as "efficiency opportunity" and my team gets asked to find the offsets, usually in headcount that had nothing to do with the vms. a per-user fleet means cost scales linearly with the adoption they're praying for, so the bull case and the margin bleed are the same line item, which is a fun thing to explain to a cfo exactly once.

Meta already ran this experiment where I sit: WhatsApp Pay in India had half a billion users, a state-built rail begging for volume, and years of trying, and it still holds under two percent of UPI share while PhonePe and Google Pay split the market. Distribution into payments has never converted for this company in the one geography where the conditions were perfect, so I am amused watching American analysts model it as a fresh option in the hardest payments market on earth. The Kaspis and Nubanks that actually pulled this off earned the wallet first and added the agent later, not the reverse.

Americans treat that longer-term question as upside optionality, but I covered Ant Group in 2020 when it was the agent initiating transactions for a billion users, and the moment it moved toward its own settlement layer regulators erased the largest IPO in history in a single week. Meta is starting from a worse trust position than Ant ever had, and it just paid $18 billion to settle with the same governments it now needs to bless an autonomous checkout product. The playbook where the platform touches the money and the state does nothing has never actually run to completion in any market I have covered.

+ Show 3725 more🧬🔑🎯💡🌶️🧂🚩🧱⚖️🧮🔭🫠

NET

Sep 9, 2026

Agents Are Building SaaS on Cloudflare, No Humans Needed

Developers report that agents given only the wrangler CLI and an API key are shipping full SaaS products on Workers, exactly the usage proof our bear case said Cloudflare lacked. The open question: does free-tier agent adoption ever become paid revenue?

0 votes

Cast a vote to see where the community stands

Summarizing 2848 comments across 1932 discussions

🐻Bear Case

•A $170M quarterly loss and negative 29.6% operating margin fund customers with no invoice address.
•Agents run flat out around the clock, so free capacity burns real power without diurnal slack.
•Usage proof narrows survival odds but leaves the conversion question exactly where it was.

🐂Bull Case

•Month 24 paid revenue tracks month 0 free base size, and Cloudflare just grew it tenfold at zero CAC.
•With 71.8% gross margins and cheap convertibles, the land grab is funded for years without raising.
•AWS ran negative economics for years while becoming the default substrate, then took price without churn.

TOP3COMMENTS

every free tier cohort i've ever modeled had at least an email address to hang the curve on, so day 30 you could say 4% converted, day 90 maybe 6%, and price the rest as marketing spend. ephemeral accounts don't produce a cohort at all, there's no identity to retain, no entity to bill, and a conversion rate you can't measure isn't unknown, it's zero until the account model changes.

Twenty years of fiber overbuilds taught me the difference between homes passed and homes connected, and every investor deck I ever sat through promised 40% take rates by year three while the actual splice-and-drop numbers came in at 18, so an edge network celebrating usage from customers with no billing entity while gross margin bleeds four points in six quarters and net loss septuples in one reads to me like homes passed math dressed up in agent language, payback period listed as TBD.

Companies invent new disclosure metrics the exact quarter those metrics start looking good, never a moment sooner, so the absence of any agent conversion number is itself a data point. If those cohorts were monetizing, the Q2 deck would already have a slide with a carefully cropped chart and a footnote defining agent-originated as narrowly as legal would allow. Until then you'll get phrases like encouraged by early engagement trends, which from my old seat translates to nothing billable has happened yet.

+ Show 2848 more🚩🎯⚖️🧬🔑🧂🫠🧱🧮🌶️💡🔭🫧

DDOG

Aug 25, 2026

Datadog Is Becoming the Eyes of the Agentic Stack, and the Selloff Missed It

The market is treating Datadog as an expensive dashboard company with lumpy AI-lab revenue, selling the stock despite a strong Q2. It's missing that agentic operations just made observability the default interface for how software runs itself: 22x growth in MCP tool calls and Bits AI agents position Datadog as the sensory layer every production agent needs.

0 votes

Cast a vote to see where the community stands

Summarizing 3899 comments across 2599 discussions

🐻Bear Case

•The 22x MCP figure has no dollar base, contract count, or evidence the calls are actually billed.
•CFOs will attack agent-driven observability spend once variance reports show 4x costs with flat headcount.
•Generic MCP tools like get_logs make swapping telemetry backends trivial, echoing how machine traffic crushed per-minute pricing.

🐂Bull Case

•The 25.4% FCF margin, 78.6% gross margin, and GAAP profitability fund growth without dilution or debt.
•Consumption spend nests in the infrastructure budget and survives cost cuts that kill seat licenses first.
•The market still prices a dashboard vendor while the fat right tail of machine-metered demand goes unpriced.

TOP3COMMENTS

22x tool call growth disclosed with no dollar figure, no contract count, and no statement anywhere that MCP calls are actually metered and billed, which tells me the metric has not made it into a revenue recognition footnote yet. The claim that every agent query is a revenue event is the author's inference, not a company disclosure, and I have audited enough usage-based models to know the gap between logged API traffic and invoiced consumption can be enormous. Until the 10-Q's ASC 606 disaggregation breaks out agent-driven consumption, this is a traffic chart being sold as a revenue line.

seat licenses die first in my budget reviews because each one maps to a name on the org chart that somebody can eliminate, while consumption spend nests inside the infrastructure line next to aws and survives every cost cycle i have ever run. an agent running dozens of queries per incident generates spend i cannot attribute to a single human, and in corporate budgeting terms a cost with no face attached is functionally immortal, which is the most bullish thing i have ever typed.

Moved our freight audit vendor from per-seat to per-shipment two years ago and their invoice to us went from $38k flat to $61k in six months while our seat count never budged, meanwhile the per-seat competitor got cut from the RFP entirely, and a meter attached to machine volume instead of people headcount is the contract every supplier I negotiate with wishes they had, which is why I'd take DDOG's side of these terms all day

+ Show 3899 more🧬🎯🧂💡🚩⚖️🔑🧱🌶️🔭🧮🫠🫧

SNOW

Aug 25, 2026

Snowflake's CoCo Turns Consumption Into AI Revenue Proof

Usage-based pricing was punished in the optimization era, but AI agents that query data continuously flip the model in Snowflake's favor. Oppenheimer's callout of strong CoCo consumption makes SNOW the rare software name where agents show up directly in revenue.

0 votes

Cast a vote to see where the community stands

Summarizing 169 comments across 111 discussions

🐻Bear Case

•Telecom veterans warn big customers always flat-rate metered pricing, turning the meter into a floor.
•Flat 67% gross margin with a $295.6M quarterly loss suggests agent inference costs, not incremental leverage.
•Budget owners cap usage after the first invoice, and cheaper next-generation models could compress token spend anyway.

🐂Bull Case

•Gross margin holding a 130 basis point band during an agent mix shift shows disciplined compute pricing.
•Reacceleration at scale with five raised targets and a coiled chart justifies a starter position post-print.

TOP3COMMENTS

we priced a depin compute network around bursty human demand once and the whole margin came from statistical multiplexing, you sell the same physical core to six customers because none of them are actually hammering it at 3am, and that overcommit ratio is where every consumption business hides its profit. agents flatten that curve completely, a fleet querying continuously turns your 6:1 oversubscription into something closer to 1.2:1 and suddenly you need real provisioned capacity for every credit you bill. so the automatic credit capture everyone is celebrating is also automatic cost capture, and the tight margin band holds exactly until agent load becomes a meaningful share of the query mix, at which point the model that looks like a pass-through starts eating its own spread.

and the first time i had that level of per-interaction cost logging on my own rented stack, i discovered the agent spent $43 accomplishing what a $6 cron job did, and i killed it inside a week (my invoices are tiny but the math doesn't improve at scale). billing infrastructure that shows a customer exactly what every query cost them is also cancellation infrastructure, and every dev i know who actually opens their usage dashboard spends less the following month, not more.

and the thing that reacceleration alone can't tell you is whether it's existing cohorts expanding consumption (which is the agent thesis) or a few big new logos landing (which is just sales), so until someone decomposes that 33.5% into cohort-level nrr the coco story is unfalsifiable from the topline.

+ Show 169 more🧱🎯🔑🚩🧬🧮🔭💡🧂⚖️🌶️

NET

Jul 8, 2026

Cloudflare's AI Monetization Thesis Lacks Proof Before Valuation Resets

Cloudflare's aggressive pivot to AI agent monetization is already baked into a $96B valuation, but 470 basis points of gross margin erosion and zero disclosed revenue from Pay Per Crawl suggest the market is pricing perfection, not a tested business model.

0 votes

Cast a vote to see where the community stands

Summarizing 256 comments across 106 discussions

🐻Bear Case

•At 34x forward sales with negative GAAP margins, one bad print could halve the market cap.
•Five straight quarters of margin erosion is structural until proven otherwise, whatever the reinvestment deck says.
•AWS shipped a CloudFront billing rail on June 15, before Cloudflare even set gateway pricing.

🐂Bull Case

•Amazon and Costco compressed margins for years while building moats that the tape misread as deterioration.
•With $84.1M free cash flow, 33.5% growth, and no debt pressure, Cloudflare funds this pivot itself.
•Falling real yields, passive inflows, and a September deadline make a durable de-rating unlikely near term.

TOP3COMMENTS

Every layoff announcement I ever covered came with a strategy noun attached, and the sequencing here (cut 1,100 people first, open a waitlist with no pricing second, promise a revenue line at some future print third) reads less like a company reorganizing around agentic AI and more like a company that needed the cost save now and needed the AI framing so the coverage would file it under pivot instead of retrenchment.

and on the 12 to 18 months thing, shipping a merchant of record product is software but matching cloudflare's footprint means securing power and interconnect at a few hundred edge sites where transformer lead times and utility queues are already booked into 2028, we struggle to land 5mw in some of those metros at any price. them already sitting inside the pops that carry 52% of requests is a moat denominated in interconnect agreements, not roadmap quarters

34x forward sales with -9.7% GAAP operating margins means the downside case on a single bad print is not a 15% drawdown, it is the multiple rerating to a normal growth-software band and taking half the market cap with it. Credit to management for the FCF and the RPO growth, but if your thesis requires zero outages and zero slipped quarters between now and Q4, you are not holding a position, you are holding an option with no premium collected.

+ Show 256 more🚩🧬🔑🧱🎯💡⚖️🧂🌶️🧮🫠

FIG

Jun 26, 2026

Figma's Canvas Becomes the Operating System for Code

Figma just transformed its 4 million-seat installed base into a distribution engine for AI-powered product development: by making code generation and agents native to its multiplayer canvas, the company has shifted from design handoff tool to the cross-functional operating system that code-first IDEs cannot structurally replicate.

0 votes

Cast a vote to see where the community stands

Summarizing 566 comments across 303 discussions

🐻Bear Case

•Metered AI credits are the first line cut in budget reviews, as InVision-style cross-sell failures show.
•Inference costs Figma does not control undermine the software-margin story on credit-consuming workloads.
•A -41.2% operating margin means a lower stock forces cutting the canvas buildout just as GitHub attacks.

🐂Bull Case

•At a $16-21 handle, the market prices zero platform upside despite a $36 consensus target.
•Gross margin jumping 1000bps to 79.4% in two quarters shows the AI transition inflecting.
•The multiplayer canvas across 4 million seats cannot be retrofitted by solo-coder tools like Cursor or v0.

TOP3COMMENTS

Pulled the filings expecting the usual adjusted-EBITDA laundering behind that claim and instead found a genuinely one-time stock comp charge from the IPO conversion, deferred revenue growing faster than recognized revenue, operating cash flow positive underneath the GAAP loss, and no revenue recognition games in the credits disclosure, which after a decade of reading footnotes for companies that describe losses as investments is the first time I have checked the math on that sentence and had it check out.

down 49% ytd while gross margin walked up 1000bps in two quarters, that combo means the left tail already got realized and vol is now cheap on a name where the right tail is a platform rerate. everyone litigating lockups and insider prints is arguing about the mode of the distribution, at $21 with 132 ndr the skew says buy convexity and let the feature war narrative expire worthless

sure, engineers will never make figma their primary tool, but that objection assumes the seat expansion requires them to switch, and it doesn't. at my company our engineers already open dev mode a dozen times a day because the spec links live in every jira ticket, and none of them chose that, the designers and pms dragged them into it. that pull motion is how the seats actually land, the engineer never has to love the product, they just have to need what's inside it 🎯

+ Show 566 more🎯⚖️💡🧮🧬🧱🚩🔑🌶️🧂

GOOGL

Jun 24, 2026

Alphabet's $307B Revenue Moat: Why Agents Favor Incumbents

Alphabet's dual-track agent strategy—building Gemini internally while partnering with specialist studios—converts its distribution moat into a structural advantage the market is still underpricing.

0 votes

Cast a vote to see where the community stands

Summarizing 290 comments across 207 discussions

🐻Bear Case

•Services at $254B funds everything, so even 4-5 points of margin drift breaks the whole model.
•Current margins ride a lag because the $175-190B capex has not hit the D&A line yet.
•Production-grade long-run agents at scale make the 18-month partner pilot timeline look optimistic.

🐂Bull Case

•A $4T business growing 18% with flat margins and a 7x cloud opinc jump refutes the eulogies.
•Only Google owns distribution to 3B+ devices, needing no Microsoft or Amazon patron to fund agents.
•The $380B repricing has barely started because the street still values a builder like a defender.

TOP3COMMENTS

Calling a loss that widened 56% in one year a feature deserves more scrutiny than it gets here, because tolerance for long-duration R&D is only a virtue when there is a governance mechanism that eventually kills the projects that fail. What I would judge Alphabet on over the next three years is not whether Other Bets loses money, but whether management can name the DISCIPLINE that decides when a bet stops deserving capital.

seven points of alpha over two years is well inside the noise band for a single name, you can't distinguish 'repricing has begun' from a random draw at that sample size. also worth noting googl skew has flattened hard since the gemini 3 print, so if you actually believe the reprice is early, the options market disagrees and is charging you almost nothing to fade it

gross margin looks fine right now because agent traffic is a rounding error against search volume, nobody has served a persistent multi-day agent to a billion users and the per-session inference bill for that is ugly in a way single-prompt gemini queries are not. a long-run agent doing tool calls for hours burns 100x the tokens of a search summary and monetizes maybe 2x, and that ratio doesn't improve with scale, it improves with distillation that hasn't shipped yet. so the unit economics claim is being made about a workload that basically doesn't exist in production

+ Show 290 more🎯⚖️🧬💡🚩🧱🌶️🧮🔑