Hanademi

Is it the end of software companies

18 slides · $3.12 · 2026-06-04 language
ENES
theme
LightDark
view
DeckTableTalk
brand
HanademiPlatzi
Is it the end of softwarecompaniesMade for Freddy Vega, by Hanademi
Is it the end of softwarecompaniesMade for Freddy Vega, by Hanademi
Anthropic annualized revenue, Jan 2024 May 2026Log scale: every gridline is ×10.Made for Freddy Vega, by HanademiSources: Nuñez, M. (2026, May 8). Anthropic says it hit a $30 billion revenue run rate. VentureBeat; Sacra. (2026).Anthropic revenue, valuation & funding.UnitUSD , annualized run-rate$10.00M$100.0M$1B$10B$100BJan 2024Dec 2024Dec 2025Feb 2026Mar 2026Apr 2026May 2026$87M$45B
From $87M to ~$45B in 28 months: the fastest revenue ramp in software history was built by selling the thing that replaces software.
Ingresos anualizados de Anthropic, ene2024 may 2026Escala logarítmica: cada línea es ×10.Made for Freddy Vega, by HanademiFuentes: Nuñez, M. (2026, May 8). Anthropic says it hit a $30 billion revenue run rate. VentureBeat; Sacra. (2026). Anthropicrevenue, valuation & funding.UnidadUSD, tasa anualizada$10,00 M$100,0 M$1 mil M$10 mil M$100 mil MJan 2024Dec 2024Dec 2025Feb 2026Mar 2026Apr 2026May 2026$87 M$45 mil M
De US$87M a ~US$45B en 28 meses: la rampa más rápida en la historia del software se construyó vendiendo aquello que reemplaza al software.
Code contributed per Anthropic engineer,pre-2025 = 1xMade for Freddy Vega, by HanademiSources: Anthropic Institute. (2026, June 4). When AI builds itself: Early signs of recursive self-improvement.Unitmultiple of pre-2025 baseline8Q2 20265.8Q1 20262.5Q4 20251.9Q3 20251.5Q2 20251.2Q1 20251Pre-2025
Inside the lab, the self-improvement loop is already turning: ~8x more code per engineer in 18 months as Claude writes Claude.
Código por ingeniero de Anthropic,pre-2025 = 1xMade for Freddy Vega, by HanademiFuentes: Anthropic Institute. (2026, June 4). When AI builds itself: Early signs of recursive self-improvement.Unidadmúltiplo de línea base pre-20258Q2 20265,8Q1 20262,5Q4 20251,9Q3 20251,5Q2 20251,2Q1 20251Pre-2025
Dentro del laboratorio, el bucle de auto-mejora ya gira: ~8x más código por ingeniero en 18 meses mientras Claude escribe a Claude.
Supabase post-money valuation,2025–2026Made for Freddy Vega, by HanademiSources: Supabase. (2026, June 4). Supabase raises $500M at $10.5B. PR Newswire; TechCrunch (2025, April 22).UnitUSDApr 2025$2BOct 2025$5BJun 2026$10.5B
Agentic infrastructure gets funded like the old SaaS winners: Supabase 5x in 14 months as agents deploy most of its databases.
Valuación post-money de Supabase,2025–2026Made for Freddy Vega, by HanademiFuentes: Supabase. (2026, June 4). Supabase raises $500M at $10.5B. PR Newswire; TechCrunch (2025, April 22).UnidadUSDApr 2025$2 mil MOct 2025$5 mil MJun 2026$10,5 mil M
La infraestructura agente se financia como los viejos ganadores SaaS: Supabase 5x en 14 meses mientras agentes despliegan la mayoría de sus bases.
Median public SaaS EV/Revenue multipleMade for Freddy Vega, by HanademiSources: Drazdou, F. (2026). SaaS Valuation Multiples: 2015–2026. Aventis Advisors; Bessemer Venture Partners. (2026). BVPNasdaq Emerging Cloud Index.UnitEV/Revenue, x19.0Dec 20207.3Feb 20253.6Feb 20263.4Mar 2026
Markets have repriced the old moat: median SaaS multiples fell >80% from 19x to 3.4x as 'build vs buy' tilts toward build.
Múltiplo EV/Ingresos mediano del SaaSpúblicoMade for Freddy Vega, by HanademiFuentes: Drazdou, F. (2026). SaaS Valuation Multiples: 2015–2026. Aventis Advisors; Bessemer Venture Partners. (2026). BVPNasdaq Emerging Cloud Index.UnidadEV/Ingresos, x19,0Dec 20207,3Feb 20253,6Feb 20263,4Mar 2026
El mercado revaluó el viejo moat: los múltiplos SaaS medianos cayeron >80% de 19x a 3,4x al inclinarse el 'construir vs. comprar' hacia construir.
AI coding output across the productionhierarchyMade for Freddy Vega, by HanademiSources: Demirer, M., Musolff, L., & Yang, A. (2026). Productivity Effects Across Generations of AI Coding Tools. SSRN; Smith,N. (2026). Noahpinion.Unit% change vs. baselineLines of code · 741%Pull requests · 65%Releases · 20%
The river of AI code narrows fast: +741% lines, +65% PRs, +20% releases — every stage downstream is the binding constraint.
Producción IA a lo largo de la jerarquíaproductivaMade for Freddy Vega, by HanademiFuentes: Demirer, M., Musolff, L., & Yang, A. (2026). Productivity Effects Across Generations of AI Coding Tools. SSRN;Smith, N. (2026). Noahpinion.Unidad% cambio vs. línea baseLíneas de código · 741 %Pull requests · 65 %Releases · 20 %
El río de código IA se angosta rápido: +741% líneas, +65% PRs, +20% releases — cada etapa río abajo es el cuello de botella.
Coding activity uplift by AI tool generationMade for Freddy Vega, by HanademiSources: Demirer, M., Musolff, L., & Yang, L. (2026). Writing Code vs. Shipping Code. NBER Working Paper 35275.Unit% lift in commits vs. baselineBaseline0%Autocomplete40%Interactive agents140%Autonomous agents180%
Each tool generation roughly doubles coding activity; releases lag far behind — proof the bottleneck is downstream, not the model.
Incremento de actividad de código porgeneración de IAMade for Freddy Vega, by HanademiFuentes: Demirer, M., Musolff, L., & Yang, L. (2026). Writing Code vs. Shipping Code. NBER Working Paper35275.Unidad% incremento en commits vs. baseLínea base0 %Autocompletado40 %Agentes interactivos140 %Agentes autónomos180 %
Cada generación de herramienta casi duplica la actividad de código; los releases quedan muy atrás — la prueba de que el cuello no es el modelo.
Where every $1 of AI coding tokens actuallygoesMade for Freddy Vega, by HanademiSources: Okoth, L. (2026, May 28). Up to 82% of AI engineering spend lost. Yahoo Finance; Smith, N. (2026). Noahpinion.Unitshare of $1 token spend0%20%40%60%80%100%Shipped productBug fixesRewrites of AI codeReview/merge delays$1 of token spend18%44%27%11%
Only 18¢ of every $1 in AI coding spend reaches a shipped feature; 82¢ is consumed by bugs, rewrites, and waiting for review.
A dónde va realmente cada US$1 en tokensIAMade for Freddy Vega, by HanademiFuentes: Okoth, L. (2026, May 28). Up to 82% of AI engineering spend lost. Yahoo Finance; Smith, N. (2026). Noahpinion.Unidadfracción de US$1 en tokens0 %20 %40 %60 %80 %100 %Producto enviadoCorrección de bugsReescrituras de código IADemoras de revisiónUS$1 en tokens18 %44 %27 %11 %
Solo 18¢ de cada US$1 en IA llega a una función enviada; 82¢ se los comen bugs, reescrituras y esperas de revisión.
At Uber, 80%+ of engineers use agentic AIand ~60% of code is AI-written yet the COOsays the link to more useful features is 'notthere yet.'Sources: Martindale, J. (2026, May 27). AI costs begin to bite. Tom's Hardware.
At Uber, 80%+ of engineers use agentic AI and ~60% of code is AI-written — yet the COO says the link to more useful features is 'not there yet.'
En Uber, 80%+ de ingenieros usa IA agente y~60% del código es IA pero su COO diceque el vínculo con mejores funciones 'aúnno existe'.Fuentes: Martindale, J. (2026, May 27). AI costs begin to bite. Tom's Hardware.
En Uber, 80%+ de ingenieros usa IA agente y ~60% del código es IA — pero su COO dice que el vínculo con mejores funciones 'aún no existe'.
METR RCT: expected vs. actual time changewith AIMade for Freddy Vega, by HanademiSources: Becker, J. et al. (2025/2026). METR developer productivity RCTs.Unit% change in task time (positive = slower)Early-2025 actual19%Early-2025 expected-24%Late-2025 original devs-18%Late-2025 new devs-4%
Experienced devs expected a 24% speedup and got a 19% slowdown — perceived productivity and measured productivity are not the same thing.
ECA METR: cambio esperado vs. real entiempo con IAMade for Freddy Vega, by HanademiFuentes: Becker, J. et al. (2025/2026). METR developer productivity RCTs.Unidad% cambio en tiempo (positivo = más lento)Inicio-2025 real19 %Inicio-2025 esperado-24 %Fin-2025 devs originales-18 %Fin-2025 devs nuevos-4 %
Los devs esperaban acelerar 24% y se ralentizaron 19% — la productividad percibida y la medida no son lo mismo.
Copilot output lift by developer tenureMade for Freddy Vega, by HanademiSources: Cui, Z. et al. (2025). The Effects of Generative AI on High-Skilled Work. MIT/SSRN.Unit% weekly output uplift39%Junior, high end27%Junior, low end26%Pooled average13%Senior, high end8%Senior, low end
Across 4,867 developers, juniors gained 27–39% and seniors only 8–13%: AI is compressing the skill premium that built the industry.
Incremento de Copilot por antigüedad deldevMade for Freddy Vega, by HanademiFuentes: Cui, Z. et al. (2025). The Effects of Generative AI on High-Skilled Work. MIT/SSRN.Unidad% incremento semanal39 %Junior, máximo27 %Junior, mínimo26 %Promedio combinado13 %Senior, máximo8 %Senior, mínimo
En 4.867 devs, los juniors ganaron 27–39% y los seniors solo 8–13%: la IA comprime la prima de habilidad que construyó la industria.
Gorgias AI ticket resolution rate by tenureMade for Freddy Vega, by HanademiSources: Montelli, A. (2026, April 28). The cheapest ticket is the one a human never touches. Gorgias ResearchLab.Unit% tickets resolved end-to-end by AIMonth 120%Month 335%Month 650%Month 965%Month 1280%
Where the workflow is owned end-to-end, AI clears 45% of tickets at the median and 70%+ at the top — the leverage shows up when judgment is encoded.
Tasa de resolución IA de tickets en Gorgiaspor antigüedadMade for Freddy Vega, by HanademiFuentes: Montelli, A. (2026, April 28). The cheapest ticket is the one a human never touches. Gorgias ResearchLab.Unidad% tickets resueltos de punta a punta por IAMes 120 %Mes 335 %Mes 650 %Mes 965 %Mes 1280 %
Cuando el flujo se posee de punta a punta, la IA resuelve 45% de tickets en la mediana y 70%+ en el tope — el apalancamiento aparece cuando el criterio se codifica.
Tellers per branch vs. urban branches,1988–2004Made for Freddy Vega, by HanademiSources: Bessen, J. (2015). Toil and Technology. IMF Finance & Development, 52(1).Unitcount / % changeTellers/branch 198820Tellers/branch 200413Urban branches Δ%43ATMs by 2015400,000
When ATMs arrived, banks did not fire tellers — they opened more branches. The first wave of automation often grows the surface area of the job.
Tellers por sucursal vs. sucursalesurbanas, 1988–2004Made for Freddy Vega, by HanademiFuentes: Bessen, J. (2015). Toil and Technology. IMF Finance & Development, 52(1).Unidadnúmero / % cambioTellers/sucursal 198820Tellers/sucursal 200413Sucursales urbanas Δ%43ATMs hacia 2015400.000
Cuando llegaron los ATMs, los bancos no despidieron tellers — abrieron más sucursales. La primera ola de automatización suele ampliar el trabajo.
U.S. bank tellers, 2004–2024Made for Freddy Vega, by HanademiSources: BLS and IMF via FRED. (2026); Autor, D. H. (2015). Why Are There Still So Many Jobs?Unitthousands of tellers0100200300317200420092012201620202024
Then mobile banking made the branch optional, and teller employment fell 42% from 2009 to 2024 — the second wave, not the first, is the one to fear.
Tellers bancarios EE.UU., 2004–2024Made for Freddy Vega, by HanademiFuentes: BLS and IMF via FRED. (2026); Autor, D. H. (2015). Why Are There Still So Many Jobs?Unidadmiles de tellers0100200300317200420092012201620202024
Luego la banca móvil hizo opcional la sucursal y los tellers cayeron 42% entre 2009 y 2024 — la segunda ola, no la primera, es la temible.
U.S. manufacturing TFP growth,electrification eraMade for Freddy Vega, by HanademiSources: David, P. A. (1989). Computer and Dynamo. Stanford CEPR.Unitannual % TFP growth1909–1919 (pre unit-drive)0.3%1919–1929 (post unit-drive)5.4%18x
Electric motors arrived in 1900 but productivity only jumped after factories were rebuilt around them — the rewiring takes a decade, not a quarter.
Crecimiento TFP manufacturero EE.UU., erade electrificaciónMade for Freddy Vega, by HanademiFuentes: David, P. A. (1989). Computer and Dynamo. Stanford CEPR.Unidad% crecimiento anual TFP1909–1919 (pre accionamiento unitario)0,3 %1919–1929 (post accionamiento unitario)5,4 %18x
Los motores eléctricos llegaron en 1900 pero la productividad solo saltó cuando las fábricas se rediseñaron — el recableado toma una década, no un trimestre.
BLS employment projections, 2024–2034Made for Freddy Vega, by HanademiSources: U.S. BLS. (2025). Accountants and Auditors; Bookkeeping, Accounting, and Auditing Clerks. Occupational Outlook Handbook.UnitjobsAccountants & auditors2024 employment2034 projected1.6K1.7KBookkeeping clerks1.1×1.6K1.5K
Spreadsheets didn't kill accounting — they killed the clerk and grew the auditor. Automation reshapes roles upward toward judgment.
Proyecciones de empleo BLS, 2024–2034Made for Freddy Vega, by HanademiFuentes: U.S. BLS. (2025). Accountants and Auditors; Bookkeeping, Accounting, and Auditing Clerks. Occupational Outlook Handbook.UnidadempleosContadores y auditoresEmpleo 2024Proyección 20341,6K1,7KAuxiliares contables1,1×1,6K1,5K
Las hojas de cálculo no mataron la contabilidad — mataron al auxiliar y crecieron al auditor. La automatización empuja los roles hacia el criterio.
Infinite automation cap by GDP shareMade for Freddy Vega, by HanademiSources: Jones, C. (2026). A.I. and Our Economic Future (v2.0). Stanford GSB.Unit% GDP uplift, 1/(1−s)100%All tasks (100%)43%Cognitive tasks (~30%)2%Software (~2% GDP)
Even free, infinite, perfect software would lift GDP by only ~2%: the value is not in the code, it's in everything code touches.
Techo de automatización infinita porparticipación en PIBMade for Freddy Vega, by HanademiFuentes: Jones, C. (2026). A.I. and Our Economic Future (v2.0). Stanford GSB.Unidad% incremento de PIB, 1/(1−s)100 %Todas las tareas (100%)43 %Tareas cognitivas (~30%)2 %Software (~2% PIB)
Incluso software gratis, infinito y perfecto subiría el PIB solo ~2%: el valor no está en el código, sino en todo lo que el código toca.
Internet adoption and daily screen timeMade for Freddy Vega, by HanademiSources: Ritchie, H. et al. (2023/2026). Our World in Data; DataReportal. (2020–2026). Digital Global Overview Reports.Unit% online / hours per day0%20%40%60%6.76.86.963%6.86.82020202120222023202420252026% world online · % onlineDaily screen time (hours) · hours/day
Two ceilings are visible at once: ~63% of humanity is online and daily screen time has plateaued near 6.8 hours — new apps now steal attention, they don't add it.
Adopción de internet y tiempo de pantalladiarioMade for Freddy Vega, by HanademiFuentes: Ritchie, H. et al. (2023/2026). Our World in Data; DataReportal. (2020–2026). Digital Global Overview Reports.Unidad% en línea / horas por día0 %20 %40 %60 %6,76,86,963 %6,86,82020202120222023202420252026% mundo en línea · % onlineTiempo pantalla diario (horas) · hours/day
Dos techos a la vez: ~63% de la humanidad está en línea y el tiempo de pantalla diario se estancó en ~6,8 horas — las nuevas apps roban atención, no la suman.
This is not the end of software companies.It is the end of companies whose only moatwas that code was hard to write.Sources: Synthesis of Aventis Advisors (2026), Jones (2026), and Supabase (2026).
This is not the end of software companies. It is the end of companies whose only moat was that code was hard to write.

La investigación detrás de esta presentación

La IA derrumba el costo de escribir código y revalúa qué vale una empresa de software. Ganan la infraestructura agente, los flujos gobernados y el criterio; pierden las empresas cuyo moat era que escribir código era difícil.

Hallazgos clave

El argumento

Esta investigación se publica en inglés y español. Read in English